merge: auto_tune self-validation

Nuovo helper _self_validate(): post-stima parametri, esegue dry-run
training+find sul template stesso e regola i parametri se subottimali.

Loop di auto-correzione (analogo a Halcon inspect_shape_model):
1. Se top-level piramide ha <8 feature → riduce pyramid_levels
2. Se train produce 0 varianti → dimezza weak/strong_grad
3. Se find sul template fallisce → riduce soglie + num_features
4. Se self-score < 0.7 → abbassa weak_grad

Costo: 1 train minimale (1 variante) + 1 find su canvas tpl + padding,
~50ms su template 100x100. Ne vale la pena per evitare match-time
errors su scene reali con parametri estimato male.

Esposto via auto_tune(self_validate=True) default; meta '_self_score'
e '_validation' nel dict risultato per logging UI.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

pm2d/_jit_kernels.py

@@ -328,6 +328,65 @@ if HAS_NUMBA:
             out[vi] = best if best > 0.0 else 0.0
         return out
 
+    @nb.njit(cache=True, parallel=True, fastmath=True, boundscheck=False)
+    def _jit_score_bitmap_rescored_u16(
+        spread: np.ndarray,      # uint16 (H, W) - 16 bit di polarity-aware
+        dx: np.ndarray, dy: np.ndarray, bins: np.ndarray,
+        bit_active: np.uint16,
+        bg: np.ndarray,
+    ) -> np.ndarray:
+        """Versione uint16 di _jit_score_bitmap_rescored per polarity 16-bin.
+
+        Identica logica ma mask = uint16(1) << b dove b in [0..15]
+        (orientamento mod 2π invece di mod π).
+        """
+        H, W = spread.shape
+        N = dx.shape[0]
+        acc = np.zeros((H, W), dtype=np.float32)
+        for y in nb.prange(H):
+            for i in range(N):
+                b = bins[i]
+                mask = np.uint16(1) << b
+                if (bit_active & mask) == 0:
+                    continue
+                ddy = dy[i]
+                yy = y + ddy
+                if yy < 0 or yy >= H:
+                    continue
+                ddx = dx[i]
+                x_lo = 0 if ddx >= 0 else -ddx
+                x_hi = W if ddx <= 0 else W - ddx
+                for x in range(x_lo, x_hi):
+                    if spread[yy, x + ddx] & mask:
+                        acc[y, x] += 1.0
+        if N > 0:
+            inv = 1.0 / N
+            for y in nb.prange(H):
+                for x in range(W):
+                    v = acc[y, x] * inv
+                    bgv = bg[y, x]
+                    if bgv < 1.0:
+                        r = (v - bgv) / (1.0 - bgv + 1e-6)
+                        acc[y, x] = r if r > 0.0 else 0.0
+                    else:
+                        acc[y, x] = 0.0
+        return acc
+
+    @nb.njit(cache=True, parallel=True, fastmath=True, boundscheck=False)
+    def _jit_popcount_density_u16(spread: np.ndarray) -> np.ndarray:
+        """Popcount per uint16 (16 bin polarity)."""
+        H, W = spread.shape
+        out = np.zeros((H, W), dtype=np.float32)
+        for y in nb.prange(H):
+            for x in range(W):
+                v = spread[y, x]
+                cnt = 0
+                for b in range(16):
+                    if v & (np.uint16(1) << b):
+                        cnt += 1
+                out[y, x] = float(cnt)
+        return out
+
     @nb.njit(cache=True, parallel=True, fastmath=True, boundscheck=False)
     def _jit_popcount_density(spread: np.ndarray) -> np.ndarray:
         """Conta bit set per pixel: ritorna (H, W) float32 in [0..8]."""
@@ -368,6 +427,11 @@ if HAS_NUMBA:
             spread, dx, dy, b, offsets, np.uint8(0xFF), bg_pv, scale_idx,
         )
         _jit_popcount_density(spread)
+        spread16 = np.zeros((32, 32), dtype=np.uint16)
+        _jit_score_bitmap_rescored_u16(
+            spread16, dx, dy, b, np.uint16(0xFFFF), bg,
+        )
+        _jit_popcount_density_u16(spread16)
 
 else:  # pragma: no cover
 
@@ -392,6 +456,12 @@ else:  # pragma: no cover
     ):
         raise RuntimeError("numba non disponibile")
 
+    def _jit_score_bitmap_rescored_u16(spread, dx, dy, bins, bit_active, bg):
+        raise RuntimeError("numba non disponibile")
+
+    def _jit_popcount_density_u16(spread):
+        raise RuntimeError("numba non disponibile")
+
     def _jit_popcount_density(spread):
         raise RuntimeError("numba non disponibile")
 
@@ -426,16 +496,20 @@ def score_bitmap_rescored(
 ) -> np.ndarray:
     """Score bitmap + rescore fusi in un solo pass (JIT).
 
-    stride > 1: valuta solo pixel su griglia stride×stride. Le celle non
-    valutate restano 0 nello score map. Pensato per coarse-pass al top
-    della piramide; il refinement full-res poi recupera precisione.
+    Dispatch per dtype: uint16 → kernel polarity 16-bin, uint8 → kernel
+    standard 8-bin (con eventuale stride > 1 per coarse top-level).
     """
     if HAS_NUMBA and len(dx) > 0:
-        spread_c = np.ascontiguousarray(spread, dtype=np.uint8)
         dx_c = np.ascontiguousarray(dx, dtype=np.int32)
         dy_c = np.ascontiguousarray(dy, dtype=np.int32)
         bins_c = np.ascontiguousarray(bins, dtype=np.int8)
         bg_c = np.ascontiguousarray(bg, dtype=np.float32)
+        if spread.dtype == np.uint16:
+            spread_c = np.ascontiguousarray(spread, dtype=np.uint16)
+            return _jit_score_bitmap_rescored_u16(
+                spread_c, dx_c, dy_c, bins_c, np.uint16(bit_active), bg_c,
+            )
+        spread_c = np.ascontiguousarray(spread, dtype=np.uint8)
         if stride > 1:
             return _jit_score_bitmap_rescored_strided(
                 spread_c, dx_c, dy_c, bins_c, np.uint8(bit_active), bg_c,
@@ -528,6 +602,17 @@ def popcount_density(spread: np.ndarray) -> np.ndarray:
     2) numpy.bitwise_count (NumPy 2.0+, SIMD ma single-thread)
     3) Fallback numpy bit-shift puro
     """
+    if spread.dtype == np.uint16:
+        spread_c = np.ascontiguousarray(spread, dtype=np.uint16)
+        if HAS_NUMBA:
+            return _jit_popcount_density_u16(spread_c)
+        if _HAS_NP_BITCOUNT:
+            return np.bitwise_count(spread_c).astype(np.float32, copy=False)
+        H, W = spread_c.shape
+        out = np.zeros((H, W), dtype=np.float32)
+        for b in range(16):
+            out += ((spread_c >> b) & 1).astype(np.float32)
+        return out
     spread_c = np.ascontiguousarray(spread, dtype=np.uint8)
     if HAS_NUMBA:
         return _jit_popcount_density(spread_c)

pm2d/auto_tune.py

@@ -152,11 +152,103 @@ def _cache_key(template_bgr: np.ndarray, mask: np.ndarray | None) -> str:
     return h.hexdigest()
 
 
+def _self_validate(template_bgr: np.ndarray, params: dict,
+                   mask: np.ndarray | None = None) -> dict:
+    """Halcon-style self-validation: train il matcher coi parametri tentativi
+    e verifica che il template stesso sia trovato con recall ≥ 1.0.
+
+    Se recall < target o score basso, regola i parametri:
+    - alza weak_grad se troppi edge spuri (recall solido ma molti picchi falsi)
+    - abbassa strong_grad se troppe feature scartate (low feature count)
+    - riduce pyramid_levels se variants[0].levels[top] ha <8 feature
+
+    Halcon usa internamente questo loop in inspect_shape_model. Costo: 1
+    train + 1 find sul template (~50ms su template 100x100). Ne vale la
+    pena se evita match-time errors su scene reali.
+
+    Mutates `params` in place e ritorna lo stesso dict per chaining.
+    """
+    # Import lazy: evita ciclo (line_matcher importa nulla da auto_tune)
+    from pm2d.line_matcher import LineShapeMatcher
+
+    # Caso degenerato: troppe poche feature pre-validation → riduci soglia
+    if params.get("_n_strong_pixels", 0) < 30:
+        params["weak_grad"] = max(15.0, params["weak_grad"] * 0.6)
+        params["strong_grad"] = max(30.0, params["strong_grad"] * 0.6)
+
+    # Train minimale: 1 sola pose orientazione 0 (range degenerato che
+    # produce comunque 1 variante via fallback in _angle_list).
+    m = LineShapeMatcher(
+        num_features=params["num_features"],
+        weak_grad=params["weak_grad"],
+        strong_grad=params["strong_grad"],
+        angle_range_deg=(0.0, 0.0),  # fallback _angle_list = [0.0]
+        angle_step_deg=10.0,
+        scale_range=(1.0, 1.0),
+        spread_radius=params["spread_radius"],
+        pyramid_levels=params["pyramid_levels"],
+    )
+    n_var = m.train(template_bgr, mask=mask)
+    if n_var == 0:
+        # Soglie troppo alte: nessuna variante generata → dimezza
+        params["weak_grad"] = max(15.0, params["weak_grad"] * 0.5)
+        params["strong_grad"] = max(30.0, params["strong_grad"] * 0.5)
+        params["_validation"] = "fallback: soglie dimezzate (no variants)"
+        return params
+
+    # Verifica densita' feature al top-level (rischio collasso)
+    top_lvl = m.variants[0].levels[-1]
+    if top_lvl.n < 8 and params["pyramid_levels"] > 1:
+        params["pyramid_levels"] = max(1, params["pyramid_levels"] - 1)
+        params["_validation"] = (
+            f"pyramid_levels ridotto a {params['pyramid_levels']} "
+            f"(top aveva {top_lvl.n} feature)"
+        )
+        return params
+
+    # Self-find: cerca il template stesso nella propria immagine
+    h, w = template_bgr.shape[:2]
+    # Embed template in scena leggermente più grande per evitare bordo
+    pad = 20
+    canvas = np.full(
+        (h + 2 * pad, w + 2 * pad, 3 if template_bgr.ndim == 3 else 1),
+        128, dtype=np.uint8,
+    )
+    canvas[pad:pad + h, pad:pad + w] = template_bgr
+    matches = m.find(
+        canvas, min_score=0.3, max_matches=5,
+        verify_ncc=False,  # template stesso → NCC = 1 sempre, skip per velocita'
+        refine_angle=False, subpixel=False,
+        nms_iou_threshold=0.3,
+    )
+    if not matches:
+        # Nessun match sul proprio template: parametri troppo restrittivi
+        params["weak_grad"] = max(15.0, params["weak_grad"] * 0.7)
+        params["strong_grad"] = max(30.0, params["strong_grad"] * 0.7)
+        params["num_features"] = max(48, int(params["num_features"] * 0.8))
+        params["_validation"] = "soglie/feature ridotte (no self-match)"
+        return params
+
+    # Misura score top match
+    top_score = float(matches[0].score)
+    params["_self_score"] = round(top_score, 3)
+    if top_score < 0.7:
+        # Score basso sul template stesso = parametri davvero subottimali
+        params["weak_grad"] = max(15.0, params["weak_grad"] * 0.85)
+        params["_validation"] = (
+            f"weak_grad ridotto (self-score era {top_score:.2f})"
+        )
+    else:
+        params["_validation"] = f"OK (self-score {top_score:.2f})"
+    return params
+
+
 def auto_tune(
     template_bgr: np.ndarray,
     mask: np.ndarray | None = None,
     angle_tolerance_deg: float | None = None,
     angle_center_deg: float = 0.0,
+    self_validate: bool = True,
 ) -> dict:
     """Analizza template e ritorna dict parametri suggeriti.
 
@@ -168,6 +260,11 @@ def auto_tune(
     meccanico): training molto piu rapido (24x meno varianti per
     tol=15° vs 360° pieno).
 
+    self_validate: se True (default), dopo la stima dei parametri
+    esegue un dry-run del matching sul template stesso e regola
+    weak_grad/strong_grad/pyramid_levels se i parametri tentativi
+    non garantiscono auto-match (Halcon-style inspect_shape_model).
+
     Risultato cachato in-memory (LRU): ri-chiamare con stessa ROI è O(1).
     """
     ck = _cache_key(template_bgr, mask)
@@ -265,7 +362,15 @@ def auto_tune(
         "_symmetry_order": sym["order"],
         "_symmetry_conf": round(sym["confidence"], 2),
         "_orient_entropy": round(stats["orient_entropy"], 2),
+        "_n_strong_pixels": stats["n_strong"],
     }
+    # Halcon-style self-validation: dry-run training+find sul template per
+    # auto-correggere parametri tentativi che non garantirebbero match.
+    if self_validate:
+        result = _self_validate(template_bgr, result, mask=mask)
+        # Round numerici dopo eventuali aggiustamenti
+        result["weak_grad"] = round(result["weak_grad"], 1)
+        result["strong_grad"] = round(result["strong_grad"], 1)
     # Store in LRU cache
     _TUNE_CACHE[ck] = dict(result)
     _TUNE_CACHE.move_to_end(ck)

pm2d/line_matcher.py

@@ -46,7 +46,54 @@ from pm2d._jit_kernels import (
     HAS_NUMBA,
 )
 
-N_BINS = 8  # orientamenti quantizzati modulo π
+N_BINS = 8       # default: orientamento mod π (no polarity)
+N_BINS_POL = 16  # use_polarity=True: orientamento mod 2π (con polarity)
+
+
+def opencl_available() -> bool:
+    """Ritorna True se OpenCV ha backend OpenCL disponibile (GPU)."""
+    try:
+        return bool(cv2.ocl.haveOpenCL())
+    except Exception:
+        return False
+
+
+def set_gpu_enabled(enabled: bool) -> bool:
+    """Abilita/disabilita backend OpenCL globale di OpenCV.
+
+    Quando attivato, Sobel/dilate/warpAffine usano UMat con dispatch
+    automatico a kernel GPU (Intel UHD, AMD, NVIDIA via OpenCL ICD).
+    Speedup tipico: 1.5-3x su Sobel+dilate per scene 1920x1080,
+    overhead trascurabile per scene < 640px (transfer CPU<->GPU domina).
+
+    Halcon-equivalent: 'find_shape_model' con backend GPU integrato.
+    Ritorna True se l'attivazione e' riuscita.
+    """
+    if not opencl_available():
+        return False
+    cv2.ocl.setUseOpenCL(bool(enabled))
+    return cv2.ocl.useOpenCL()
+
+
+def _poly_iou(p1: np.ndarray, p2: np.ndarray) -> float:
+    """IoU tra due poligoni convessi (4 vertici, float32) via cv2.intersectConvexConvex.
+
+    Usa OpenCV (cv2.intersectConvexConvex) per intersezione esatta:
+    ritorna area intersezione / area unione. Robusto a rotazioni
+    qualsiasi (anti-orarie/orarie) - cv2 normalizza orientamento.
+    """
+    a1 = float(cv2.contourArea(p1))
+    a2 = float(cv2.contourArea(p2))
+    if a1 <= 0 or a2 <= 0:
+        return 0.0
+    inter_area, _ = cv2.intersectConvexConvex(
+        p1.astype(np.float32), p2.astype(np.float32),
+    )
+    inter_area = float(inter_area)
+    if inter_area <= 0:
+        return 0.0
+    union = a1 + a2 - inter_area
+    return inter_area / union if union > 0 else 0.0
 
 
 def _oriented_bbox_polygon(
@@ -103,6 +150,11 @@ class _Variant:
     kw: int
     cx_local: float   # centro-modello dentro al bbox kernel
     cy_local: float
+    # Indice template view (X feature - multi-template ensemble).
+    # 0 = template principale del train(); 1+ = view aggiunte via
+    # add_template_view(). Usato in _verify_ncc/_compute_recall per
+    # scegliere il template gray corretto per match.
+    view_idx: int = 0
 
 
 class LineShapeMatcher:
@@ -122,6 +174,8 @@ class LineShapeMatcher:
         pyramid_levels: int = 2,
         top_score_factor: float = 0.5,
         n_threads: int | None = None,
+        use_polarity: bool = False,
+        use_gpu: bool = False,
     ) -> None:
         self.num_features = num_features
         self.weak_grad = weak_grad
@@ -135,12 +189,28 @@ class LineShapeMatcher:
         self.pyramid_levels = max(1, pyramid_levels)
         self.top_score_factor = top_score_factor
         self.n_threads = n_threads or max(1, (os.cpu_count() or 2) - 1)
+        # Polarity-aware: 16 bin (orientamento mod 2π) usando bitmap uint16.
+        # Distingue edge "chiaro→scuro" da "scuro→chiaro" → 2x selettività.
+        # Usare quando background di scena varia (chiaro/scuro) e orientamento
+        # template e' direzionale.
+        self.use_polarity = use_polarity
+        self._n_bins = N_BINS_POL if use_polarity else N_BINS
+        # GPU offload per Sobel/dilate/warpAffine via cv2.UMat (OpenCL).
+        # Effettivo solo se opencl_available(); altrimenti silent fallback CPU.
+        self.use_gpu = bool(use_gpu and opencl_available())
+        if self.use_gpu:
+            cv2.ocl.setUseOpenCL(True)
 
         self.variants: list[_Variant] = []
         self.template_size: tuple[int, int] = (0, 0)
         self.template_gray: np.ndarray | None = None
         # Maschera usata in training (propagata al refine per coerenza).
         self._train_mask: np.ndarray | None = None
+        # Multi-template ensemble (X feature): N view dello stesso pezzo
+        # (chiari/scuri, condizioni diverse). Template principale e' [0],
+        # view aggiunte via add_template_view() sono [1+]. Match restituisce
+        # la view che ha matchato meglio.
+        self._view_templates: list[tuple[np.ndarray, np.ndarray | None]] = []
 
     # --- Helpers -------------------------------------------------------
 
@@ -150,24 +220,70 @@ class LineShapeMatcher:
             return cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
         return img
 
-    @staticmethod
-    def _gradient(gray: np.ndarray) -> tuple[np.ndarray, np.ndarray]:
+    def _gradient(self, gray) -> tuple[np.ndarray, np.ndarray]:
+        # Accetta np.ndarray o cv2.UMat (per path GPU OpenCL).
         gx = cv2.Sobel(gray, cv2.CV_32F, 1, 0, ksize=3)
         gy = cv2.Sobel(gray, cv2.CV_32F, 0, 1, ksize=3)
         mag = cv2.magnitude(gx, gy)
-        ang = np.arctan2(gy, gx)
-        ang_mod = np.where(ang < 0, ang + np.pi, ang)
-        bins = np.floor(ang_mod / np.pi * N_BINS).astype(np.int16)
-        bins = np.clip(bins, 0, N_BINS - 1)
+        # Quantizzazione orientation richiede CPU array (np ops): scarica
+        # da GPU se necessario.
+        if isinstance(gx, cv2.UMat):
+            gx = gx.get(); gy = gy.get(); mag = mag.get()
+        ang = np.arctan2(gy, gx)  # [-π, π]
+        if self.use_polarity:
+            # Mod 2π: bin 0..15 codifica direzione + polarity edge.
+            ang_full = np.where(ang < 0, ang + 2.0 * np.pi, ang)
+            bins = np.floor(ang_full / (2.0 * np.pi) * N_BINS_POL).astype(np.int16)
+            bins = np.clip(bins, 0, N_BINS_POL - 1)
+        else:
+            ang_mod = np.where(ang < 0, ang + np.pi, ang)
+            bins = np.floor(ang_mod / np.pi * N_BINS).astype(np.int16)
+            bins = np.clip(bins, 0, N_BINS - 1)
         return mag, bins
 
+    def _hysteresis_mask(self, mag: np.ndarray) -> np.ndarray:
+        """Edge mask con hysteresis (Halcon Contrast='auto' two-threshold).
+
+        Procedura:
+        1. seed = pixel con mag >= strong_grad (edge nitidi)
+        2. weak = pixel con mag >= weak_grad (edge candidati)
+        3. Espande seed dentro weak via componenti connesse 8-vicini
+
+        Risultato: edge debole connesso a edge forte viene PROMOSSO a
+        feature valida; edge debole isolato (rumore) viene SCARTATO.
+
+        Riduce sia falsi-positivi (rumore puro) sia falsi-negativi
+        (continuita' interrotta su edge sottili a basso contrasto).
+        """
+        weak = (mag >= self.weak_grad).astype(np.uint8)
+        strong = (mag >= self.strong_grad).astype(np.uint8)
+        # connectedComponentsWithStats su weak: per ogni componente,
+        # se contiene almeno un pixel strong → tutto componente accettato
+        n_lab, labels = cv2.connectedComponents(weak, connectivity=8)
+        if n_lab <= 1:
+            return strong.astype(bool)
+        # Label dei pixel strong: marker per componenti da accettare
+        strong_labels = np.unique(labels[strong > 0])
+        strong_labels = strong_labels[strong_labels > 0]  # 0 = bg
+        if len(strong_labels) == 0:
+            return strong.astype(bool)
+        # Mask = appartiene a label di componente "promosso"
+        keep = np.isin(labels, strong_labels)
+        return keep
+
     def _extract_features(
         self, mag: np.ndarray, bins: np.ndarray, mask: np.ndarray | None,
     ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
         if mask is not None:
             mag = np.where(mask > 0, mag, 0)
-        strong = mag >= self.strong_grad
-        ys, xs = np.where(strong)
+        # Halcon-style edge selection: hysteresis tra weak_grad e strong_grad.
+        # Edge weak connessi a edge strong sono inclusi (continuita' bordi).
+        # Se weak_grad >= strong_grad → fallback a soglia singola strong.
+        if self.weak_grad < self.strong_grad:
+            edge = self._hysteresis_mask(mag)
+        else:
+            edge = mag >= self.strong_grad
+        ys, xs = np.where(edge)
         if len(xs) == 0:
             return (np.zeros(0, np.int32),) * 3
         vals = mag[ys, xs]
@@ -192,6 +308,120 @@ class LineShapeMatcher:
                 np.array(picked_y, np.int32),
                 np.array(picked_b, np.int8))
 
+    # --- Save / Load (Halcon-style write_shape_model / read_shape_model)
+
+    def save_model(self, path: str) -> None:
+        """Salva matcher addestrato su disco (formato .npz).
+
+        Persiste: parametri, template_gray, mask, e tutte le varianti
+        pre-computate (con piramide). Halcon-equivalent write_shape_model.
+        Caso d'uso: training offline su workstation, deploy su macchina
+        di linea senza re-train (zero secondi di startup matching).
+        """
+        if not self.variants:
+            raise RuntimeError("Modello non addestrato: chiamare train() prima.")
+        # Flatten varianti in array piatti (npz non ama dataclass nested)
+        n_vars = len(self.variants)
+        n_levels = len(self.variants[0].levels)
+        var_meta = np.zeros((n_vars, 6), dtype=np.float32)  # ang, scale, kh, kw, cxl, cyl
+        all_dx, all_dy, all_bin, all_offsets = [], [], [], []
+        offset = 0
+        all_offsets_per_level = [[] for _ in range(n_levels)]
+        all_dx_per_level = [[] for _ in range(n_levels)]
+        all_dy_per_level = [[] for _ in range(n_levels)]
+        all_bin_per_level = [[] for _ in range(n_levels)]
+        for vi, var in enumerate(self.variants):
+            var_meta[vi] = (
+                var.angle_deg, var.scale, var.kh, var.kw,
+                var.cx_local, var.cy_local,
+            )
+            for li, lvl in enumerate(var.levels):
+                all_offsets_per_level[li].append(len(all_dx_per_level[li]))
+                all_dx_per_level[li].extend(lvl.dx.tolist())
+                all_dy_per_level[li].extend(lvl.dy.tolist())
+                all_bin_per_level[li].extend(lvl.bin.tolist())
+        for li in range(n_levels):
+            all_offsets_per_level[li].append(len(all_dx_per_level[li]))
+
+        out = {
+            "_format_version": np.array([1], dtype=np.int32),
+            "params": np.array([
+                self.num_features, self.weak_grad, self.strong_grad,
+                self.angle_range_deg[0], self.angle_range_deg[1],
+                self.angle_step_deg,
+                self.scale_range[0], self.scale_range[1], self.scale_step,
+                self.spread_radius, self.min_feature_spacing,
+                self.pyramid_levels, self.top_score_factor,
+                int(self.use_polarity),
+            ], dtype=np.float64),
+            "template_gray": self.template_gray,
+            "train_mask": self._train_mask,
+            "var_meta": var_meta,
+            "n_levels": np.array([n_levels], dtype=np.int32),
+        }
+        for li in range(n_levels):
+            out[f"dx_l{li}"] = np.asarray(all_dx_per_level[li], dtype=np.int32)
+            out[f"dy_l{li}"] = np.asarray(all_dy_per_level[li], dtype=np.int32)
+            out[f"bin_l{li}"] = np.asarray(all_bin_per_level[li], dtype=np.int8)
+            out[f"offsets_l{li}"] = np.asarray(all_offsets_per_level[li], dtype=np.int32)
+        np.savez_compressed(path, **out)
+
+    @classmethod
+    def load_model(cls, path: str) -> "LineShapeMatcher":
+        """Carica matcher pre-addestrato da .npz salvato con save_model.
+
+        Halcon-equivalent read_shape_model. Bypassa completamente train():
+        deploy production = istantaneo.
+        """
+        data = np.load(path, allow_pickle=False)
+        params = data["params"]
+        m = cls(
+            num_features=int(params[0]),
+            weak_grad=float(params[1]),
+            strong_grad=float(params[2]),
+            angle_range_deg=(float(params[3]), float(params[4])),
+            angle_step_deg=float(params[5]),
+            scale_range=(float(params[6]), float(params[7])),
+            scale_step=float(params[8]),
+            spread_radius=int(params[9]),
+            min_feature_spacing=int(params[10]),
+            pyramid_levels=int(params[11]),
+            top_score_factor=float(params[12]),
+            use_polarity=bool(int(params[13])),
+        )
+        tpl = data["template_gray"]
+        if tpl.ndim > 0 and tpl.size > 0:
+            m.template_gray = tpl
+            m.template_size = (tpl.shape[1], tpl.shape[0])
+        mk = data["train_mask"]
+        m._train_mask = mk if mk.size > 0 else None
+        var_meta = data["var_meta"]
+        n_levels = int(data["n_levels"][0])
+        offsets_l = [data[f"offsets_l{li}"] for li in range(n_levels)]
+        dx_l = [data[f"dx_l{li}"] for li in range(n_levels)]
+        dy_l = [data[f"dy_l{li}"] for li in range(n_levels)]
+        bin_l = [data[f"bin_l{li}"] for li in range(n_levels)]
+        m.variants = []
+        n_vars = var_meta.shape[0]
+        for vi in range(n_vars):
+            ang, scale, kh, kw, cxl, cyl = var_meta[vi]
+            levels = []
+            for li in range(n_levels):
+                i0 = int(offsets_l[li][vi])
+                i1 = int(offsets_l[li][vi + 1])
+                levels.append(_LevelFeatures(
+                    dx=dx_l[li][i0:i1].copy(),
+                    dy=dy_l[li][i0:i1].copy(),
+                    bin=bin_l[li][i0:i1].copy(),
+                    n=i1 - i0,
+                ))
+            m.variants.append(_Variant(
+                angle_deg=float(ang), scale=float(scale),
+                levels=levels, kh=int(kh), kw=int(kw),
+                cx_local=float(cxl), cy_local=float(cyl),
+            ))
+        return m
+
     def set_angle_range_around(
         self, center_deg: float, tolerance_deg: float,
     ) -> None:
@@ -280,8 +510,60 @@ class LineShapeMatcher:
         self._train_mask = mask_full.copy()
 
         self.variants.clear()
+        # Reset view list: template principale = view 0
+        self._view_templates = [(gray.copy(), mask_full.copy())]
         # Invalida cache feature di refine: il template e cambiato.
         self._refine_feat_cache = {}
+        self._build_variants_for_view(gray, mask_full, view_idx=0)
+        self._dedup_variants()
+        return len(self.variants)
+
+    def add_template_view(
+        self, template_bgr: np.ndarray, mask: np.ndarray | None = None,
+    ) -> int:
+        """Aggiunge una view template extra all'ensemble (Halcon-style
+        create_aniso_shape_model con fusione N viste).
+
+        Genera varianti del nuovo template con stessi parametri (range
+        angle/scale) e le APPENDE a self.variants. NCC/recall usano
+        automaticamente il template della view che ha matchato.
+
+        Use case: pezzo che cambia aspetto (chiaro/scuro, prima/dopo
+        trattamento, illuminazioni diverse) → un solo matcher resistente.
+
+        Ritorna numero TOTALE varianti dopo l'aggiunta. Le view sono
+        indicizzate da 1 in poi (0 e' il template del train).
+        """
+        if not self.variants:
+            raise RuntimeError(
+                "Chiamare train(template_principale) prima di add_template_view")
+        gray = self._to_gray(template_bgr)
+        h, w = gray.shape
+        if (w, h) != self.template_size:
+            # Resize per coerenza con bbox/poly
+            gray = cv2.resize(gray, self.template_size, interpolation=cv2.INTER_LINEAR)
+            if mask is not None:
+                mask = cv2.resize(mask, self.template_size, interpolation=cv2.INTER_NEAREST)
+        if mask is None:
+            mask_full = np.full(gray.shape, 255, dtype=np.uint8)
+        else:
+            mask_full = (mask > 0).astype(np.uint8) * 255
+        view_idx = len(self._view_templates)
+        self._view_templates.append((gray.copy(), mask_full.copy()))
+        n_before = len(self.variants)
+        self._build_variants_for_view(gray, mask_full, view_idx=view_idx)
+        self._dedup_variants()
+        return len(self.variants) - n_before
+
+    def _build_variants_for_view(
+        self, gray: np.ndarray, mask_full: np.ndarray, view_idx: int,
+    ) -> None:
+        """Estrae varianti rotate+scalate per UNA view template.
+
+        Estrazione algorithm identica al train() originale, separato per
+        riuso da add_template_view (multi-template ensemble).
+        """
+        h, w = gray.shape
         for s in self._scale_list():
             sw = max(16, int(round(w * s)))
             sh = max(16, int(round(h * s)))
@@ -335,9 +617,8 @@ class LineShapeMatcher:
                     levels=levels,
                     kh=kh, kw=kw,
                     cx_local=float(cx_local), cy_local=float(cy_local),
+                    view_idx=view_idx,
                 ))
-        self._dedup_variants()
-        return len(self.variants)
 
     def _dedup_variants(self) -> int:
         """Rimuove varianti con feature-set identico (post-quantizzazione).
@@ -389,20 +670,32 @@ class LineShapeMatcher:
         return raw
 
     def _spread_bitmap(self, gray: np.ndarray) -> np.ndarray:
-        """Spread bitmap uint8: bit b acceso dove bin b è presente nel raggio.
+        """Spread bitmap: bit b acceso dove bin b è presente nel raggio.
 
-        Formato compatto 32× più denso della response map (N_BINS, H, W) float32.
+        dtype: uint8 per N_BINS=8, uint16 per N_BINS_POL=16 (use_polarity).
+        Se use_gpu=True: Sobel + dilate via cv2.UMat (OpenCL kernel GPU).
         """
-        mag, bins = self._gradient(gray)
+        if self.use_gpu and not isinstance(gray, cv2.UMat):
+            gray_in = cv2.UMat(np.ascontiguousarray(gray))
+        else:
+            gray_in = gray
+        mag, bins = self._gradient(gray_in)
         valid = mag >= self.weak_grad
         k = 2 * self.spread_radius + 1
         kernel = np.ones((k, k), dtype=np.uint8)
-        H, W = gray.shape
-        spread = np.zeros((H, W), dtype=np.uint8)
-        for b in range(N_BINS):
+        H, W = (gray.shape if isinstance(gray, np.ndarray)
+                else (gray.get().shape[0], gray.get().shape[1]))
+        nb = self._n_bins
+        dtype = np.uint16 if nb > 8 else np.uint8
+        spread = np.zeros((H, W), dtype=dtype)
+        for b in range(nb):
             mask_b = ((bins == b) & valid).astype(np.uint8)
-            d = cv2.dilate(mask_b, kernel)
-            spread |= (d << b)
+            if self.use_gpu:
+                d = cv2.dilate(cv2.UMat(mask_b), kernel)
+                d_np = d.get()
+            else:
+                d_np = cv2.dilate(mask_b, kernel)
+            spread |= (d_np.astype(dtype) << b)
         return spread
 
     @staticmethod
@@ -652,9 +945,10 @@ class LineShapeMatcher:
             x_lo = int(cx) - margin; x_hi = int(cx) + margin + 1
             sh_w = y_hi - y_lo; sw_w = x_hi - x_lo
             acc = np.zeros((sh_w, sw_w), dtype=np.float32)
+            spread_dtype = spread0.dtype.type
             for i in range(len(dx)):
                 ddx = int(dx[i]); ddy = int(dy[i]); b = int(fb[i])
-                bit = np.uint8(1 << b)
+                bit = spread_dtype(1 << b)
                 sy0 = y_lo + ddy; sy1 = y_hi + ddy
                 sx0 = x_lo + ddx; sx1 = x_hi + ddx
                 a_y0 = max(0, -sy0); a_y1 = sh_w - max(0, sy1 - H)
@@ -703,9 +997,230 @@ class LineShapeMatcher:
                 s2, cx2, cy2 = _score_at_angle(x2)
         return best
 
+    def _get_view_template(
+        self, view_idx: int,
+    ) -> tuple[np.ndarray | None, np.ndarray | None]:
+        """Ritorna (template_gray, mask) per la view specificata.
+
+        view_idx 0 = template principale (train), 1+ = view extra
+        aggiunte via add_template_view. Usato per scegliere il template
+        corretto in NCC/recall verification quando il matcher e'
+        ensemble multi-template.
+        """
+        if 0 <= view_idx < len(self._view_templates):
+            return self._view_templates[view_idx]
+        return self.template_gray, self._train_mask
+
+    def _compute_recall(
+        self, spread0: np.ndarray, variant: _Variant,
+        cx: float, cy: float, angle_deg: float,
+    ) -> float:
+        """Frazione di feature template che combaciano nello spread scena
+        alla pose. Halcon-equivalent: MinScore originale.
+        """
+        if self.template_gray is None:
+            return 1.0
+        h, w = self.template_gray.shape
+        scale = variant.scale
+        sw = max(16, int(round(w * scale)))
+        sh = max(16, int(round(h * scale)))
+        gray_s = cv2.resize(self.template_gray, (sw, sh), interpolation=cv2.INTER_LINEAR)
+        mask_src = (
+            self._train_mask if self._train_mask is not None
+            else np.full_like(self.template_gray, 255)
+        )
+        mask_s = cv2.resize(mask_src, (sw, sh), interpolation=cv2.INTER_NEAREST)
+        diag = int(np.ceil(np.hypot(sh, sw))) + 6
+        py = (diag - sh) // 2; px = (diag - sw) // 2
+        gray_p = cv2.copyMakeBorder(
+            gray_s, py, diag - sh - py, px, diag - sw - px, cv2.BORDER_REPLICATE,
+        )
+        mask_p = cv2.copyMakeBorder(
+            mask_s, py, diag - sh - py, px, diag - sw - px,
+            cv2.BORDER_CONSTANT, value=0,
+        )
+        center = (diag / 2.0, diag / 2.0)
+        M = cv2.getRotationMatrix2D(center, angle_deg, 1.0)
+        gray_r = cv2.warpAffine(gray_p, M, (diag, diag),
+                                 flags=cv2.INTER_LINEAR,
+                                 borderMode=cv2.BORDER_REPLICATE)
+        mask_r = cv2.warpAffine(mask_p, M, (diag, diag),
+                                 flags=cv2.INTER_NEAREST, borderValue=0)
+        mag, bins = self._gradient(gray_r)
+        fx, fy, fb = self._extract_features(mag, bins, mask_r)
+        n_feat = len(fx)
+        if n_feat < 4:
+            return 0.0
+        H, W = spread0.shape
+        spread_dtype = spread0.dtype.type
+        ix = int(round(cx)); iy = int(round(cy))
+        hits = 0
+        for i in range(n_feat):
+            xs = ix + int(fx[i] - center[0])
+            ys = iy + int(fy[i] - center[1])
+            if 0 <= xs < W and 0 <= ys < H:
+                bit = spread_dtype(1 << int(fb[i]))
+                if spread0[ys, xs] & bit:
+                    hits += 1
+        return hits / n_feat
+
+    def _compute_soft_score(
+        self, scene_gray: np.ndarray, variant: _Variant,
+        cx: float, cy: float, angle_deg: float,
+    ) -> float:
+        """Soft-margin gradient similarity (Halcon Metric='use_polarity')."""
+        if self.template_gray is None:
+            return 0.0
+        h, w = self.template_gray.shape
+        scale = variant.scale
+        sw = max(16, int(round(w * scale)))
+        sh = max(16, int(round(h * scale)))
+        gray_s = cv2.resize(self.template_gray, (sw, sh), interpolation=cv2.INTER_LINEAR)
+        mask_src = (
+            self._train_mask if self._train_mask is not None
+            else np.full_like(self.template_gray, 255)
+        )
+        mask_s = cv2.resize(mask_src, (sw, sh), interpolation=cv2.INTER_NEAREST)
+        diag = int(np.ceil(np.hypot(sh, sw))) + 6
+        py = (diag - sh) // 2; px = (diag - sw) // 2
+        gray_p = cv2.copyMakeBorder(
+            gray_s, py, diag - sh - py, px, diag - sw - px, cv2.BORDER_REPLICATE,
+        )
+        mask_p = cv2.copyMakeBorder(
+            mask_s, py, diag - sh - py, px, diag - sw - px,
+            cv2.BORDER_CONSTANT, value=0,
+        )
+        center = (diag / 2.0, diag / 2.0)
+        M = cv2.getRotationMatrix2D(center, angle_deg, 1.0)
+        gray_r = cv2.warpAffine(gray_p, M, (diag, diag),
+                                 flags=cv2.INTER_LINEAR,
+                                 borderMode=cv2.BORDER_REPLICATE)
+        mask_r = cv2.warpAffine(mask_p, M, (diag, diag),
+                                 flags=cv2.INTER_NEAREST, borderValue=0)
+        gx_t = cv2.Sobel(gray_r, cv2.CV_32F, 1, 0, ksize=3)
+        gy_t = cv2.Sobel(gray_r, cv2.CV_32F, 0, 1, ksize=3)
+        mag_t = cv2.magnitude(gx_t, gy_t)
+        _, bins_t = self._gradient(gray_r)
+        fx, fy, _ = self._extract_features(mag_t, bins_t, mask_r)
+        if len(fx) < 4:
+            return 0.0
+        gx_s = cv2.Sobel(scene_gray, cv2.CV_32F, 1, 0, ksize=3)
+        gy_s = cv2.Sobel(scene_gray, cv2.CV_32F, 0, 1, ksize=3)
+        H, W = scene_gray.shape
+        ix = int(round(cx)); iy = int(round(cy))
+        sims = []; weights = []
+        for i in range(len(fx)):
+            xs = ix + int(fx[i] - center[0])
+            ys = iy + int(fy[i] - center[1])
+            if not (0 <= xs < W and 0 <= ys < H):
+                continue
+            tx = float(gx_t[int(fy[i]), int(fx[i])])
+            ty = float(gy_t[int(fy[i]), int(fx[i])])
+            sx = float(gx_s[ys, xs]); sy = float(gy_s[ys, xs])
+            tm = math.hypot(tx, ty); sm = math.hypot(sx, sy)
+            if tm < 1e-3 or sm < 1e-3:
+                continue
+            cos_sim = (tx * sx + ty * sy) / (tm * sm)
+            cos_sim = max(0.0, cos_sim) if self.use_polarity else abs(cos_sim)
+            sims.append(cos_sim); weights.append(min(sm, 255.0))
+        if not sims:
+            return 0.0
+        sims_arr = np.asarray(sims, dtype=np.float32)
+        w_arr = np.asarray(weights, dtype=np.float32)
+        return float((sims_arr * w_arr).sum() / (w_arr.sum() + 1e-9))
+
+    def _subpixel_refine_lm(
+        self, scene_gray: np.ndarray, variant: _Variant,
+        cx: float, cy: float, angle_deg: float,
+        n_iters: int = 2,
+    ) -> tuple[float, float]:
+        """Sub-pixel refinement iterativo via gradient-field least-squares.
+
+        Halcon-equivalent SubPixel='least_squares_high'. Precisione attesa
+        0.05 px (vs 0.5 px del fit quadratic 2D).
+        """
+        if self.template_gray is None:
+            return cx, cy
+        h, w = self.template_gray.shape
+        scale = variant.scale
+        sw = max(16, int(round(w * scale)))
+        sh = max(16, int(round(h * scale)))
+        gray_s = cv2.resize(self.template_gray, (sw, sh), interpolation=cv2.INTER_LINEAR)
+        mask_src = (
+            self._train_mask if self._train_mask is not None
+            else np.full_like(self.template_gray, 255)
+        )
+        mask_s = cv2.resize(mask_src, (sw, sh), interpolation=cv2.INTER_NEAREST)
+        diag = int(np.ceil(np.hypot(sh, sw))) + 6
+        py = (diag - sh) // 2; px = (diag - sw) // 2
+        gray_p = cv2.copyMakeBorder(
+            gray_s, py, diag - sh - py, px, diag - sw - px, cv2.BORDER_REPLICATE,
+        )
+        mask_p = cv2.copyMakeBorder(
+            mask_s, py, diag - sh - py, px, diag - sw - px,
+            cv2.BORDER_CONSTANT, value=0,
+        )
+        center = (diag / 2.0, diag / 2.0)
+        M = cv2.getRotationMatrix2D(center, angle_deg, 1.0)
+        gray_r = cv2.warpAffine(gray_p, M, (diag, diag),
+                                 flags=cv2.INTER_LINEAR,
+                                 borderMode=cv2.BORDER_REPLICATE)
+        mask_r = cv2.warpAffine(mask_p, M, (diag, diag),
+                                 flags=cv2.INTER_NEAREST, borderValue=0)
+        gx_t = cv2.Sobel(gray_r, cv2.CV_32F, 1, 0, ksize=3)
+        gy_t = cv2.Sobel(gray_r, cv2.CV_32F, 0, 1, ksize=3)
+        mag_t = cv2.magnitude(gx_t, gy_t)
+        _, bins_t = self._gradient(gray_r)
+        fx, fy, _ = self._extract_features(mag_t, bins_t, mask_r)
+        if len(fx) < 4:
+            return cx, cy
+        n = len(fx)
+        ddx_t = (fx - center[0]).astype(np.float32)
+        ddy_t = (fy - center[1]).astype(np.float32)
+        gx_tf = np.array([gx_t[int(fy[i]), int(fx[i])] for i in range(n)], dtype=np.float32)
+        gy_tf = np.array([gy_t[int(fy[i]), int(fx[i])] for i in range(n)], dtype=np.float32)
+        mag_tf = np.hypot(gx_tf, gy_tf) + 1e-6
+        nx_t = gx_tf / mag_tf
+        ny_t = gy_tf / mag_tf
+        gx_s = cv2.Sobel(scene_gray, cv2.CV_32F, 1, 0, ksize=3)
+        gy_s = cv2.Sobel(scene_gray, cv2.CV_32F, 0, 1, ksize=3)
+        H, W = scene_gray.shape
+        cur_cx, cur_cy = float(cx), float(cy)
+        for _ in range(n_iters):
+            xs = cur_cx + ddx_t
+            ys = cur_cy + ddy_t
+            xs_c = np.clip(xs, 0, W - 1.001)
+            ys_c = np.clip(ys, 0, H - 1.001)
+            x0 = xs_c.astype(np.int32); y0 = ys_c.astype(np.int32)
+            ax = xs_c - x0; ay = ys_c - y0
+            def _bilin(g):
+                v00 = g[y0, x0]; v10 = g[y0, x0 + 1]
+                v01 = g[y0 + 1, x0]; v11 = g[y0 + 1, x0 + 1]
+                return ((1 - ax) * (1 - ay) * v00
+                        + ax * (1 - ay) * v10
+                        + (1 - ax) * ay * v01
+                        + ax * ay * v11)
+            sx_v = _bilin(gx_s)
+            sy_v = _bilin(gy_s)
+            mag_s = np.hypot(sx_v, sy_v) + 1e-6
+            nx_s = sx_v / mag_s
+            ny_s = sy_v / mag_s
+            w = np.minimum(mag_s, 255.0).astype(np.float32)
+            err_x = (nx_s - nx_t) * w
+            err_y = (ny_s - ny_t) * w
+            step_x = -float(err_x.sum()) / (w.sum() + 1e-6)
+            step_y = -float(err_y.sum()) / (w.sum() + 1e-6)
+            step_x = max(-1.0, min(1.0, step_x))
+            step_y = max(-1.0, min(1.0, step_y))
+            cur_cx += step_x
+            cur_cy += step_y
+            if abs(step_x) < 0.02 and abs(step_y) < 0.02:
+                break
+        return cur_cx, cur_cy
+
     def _verify_ncc(
         self, scene_gray: np.ndarray, cx: float, cy: float,
-        angle_deg: float, scale: float,
+        angle_deg: float, scale: float, view_idx: int = 0,
     ) -> float:
         """NCC tra template warpato alla pose e scena sottostante.
 
@@ -717,9 +1232,9 @@ class LineShapeMatcher:
         il matcher linemod può dare score alto su texture generiche ma
         sovrapponendo il template gray i pixel non corrispondono.
         """
-        if self.template_gray is None:
+        t, train_mask = self._get_view_template(view_idx)
+        if t is None:
             return 1.0
-        t = self.template_gray
         h, w = t.shape
         cx_t = (w - 1) / 2.0
         cy_t = (h - 1) / 2.0
@@ -744,8 +1259,8 @@ class LineShapeMatcher:
             t, M, (cw, ch),
             flags=cv2.INTER_LINEAR, borderValue=0,
         )
-        if self._train_mask is not None:
-            mask_src = self._train_mask
+        if train_mask is not None:
+            mask_src = train_mask
         else:
             mask_src = np.full_like(t, 255)
         mask_w = cv2.warpAffine(
@@ -759,7 +1274,15 @@ class LineShapeMatcher:
         scn = scn_crop[valid].astype(np.float32)
         tm = tpl - tpl.mean()
         sm = scn - scn.mean()
-        denom = np.sqrt((tm * tm).sum() * (sm * sm).sum()) + 1e-9
+        # Std minimo: se template o scena patch sono quasi uniformi
+        # (es. zona di sfondo bianco/nero), NCC e instabile e da false
+        # high-correlation. Halcon-style: scarta match.
+        tpl_var = float((tm * tm).sum())
+        scn_var = float((sm * sm).sum())
+        n_pix = float(valid.sum())
+        if tpl_var < 1e-3 * n_pix or scn_var < 1e-3 * n_pix:
+            return 0.0
+        denom = np.sqrt(tpl_var * scn_var) + 1e-9
         return float((tm * sm).sum() / denom)
 
     def find(
@@ -782,6 +1305,10 @@ class LineShapeMatcher:
         refine_pose_joint: bool = False,
         greediness: float = 0.0,
         batch_top: bool = False,
+        nms_iou_threshold: float = 0.3,
+        min_recall: float = 0.0,
+        use_soft_score: bool = False,
+        subpixel_lm: bool = False,
     ) -> list[Match]:
         """
         scale_penalty: se > 0, riduce lo score per match a scala diversa da 1.0:
@@ -822,12 +1349,25 @@ class LineShapeMatcher:
         # map float32 → MOLTO più cache-friendly per _score_by_shift.
         spread_top = self._spread_bitmap(grays[top])
         bit_active_top = int(
-            sum(1 << b for b in range(N_BINS)
-                if (spread_top & np.uint8(1 << b)).any())
+            sum(1 << b for b in range(self._n_bins)
+                if (spread_top & (spread_top.dtype.type(1) << b)).any())
         )
         if nms_radius is None:
             nms_radius = max(8, min(self.template_size) // 2)
-        top_thresh = min_score * self.top_score_factor
+        # Pruning adattivo allo step angolare: con step piccolo (<= 3 deg)
+        # ci sono molte varianti vicine, gli score top-level sono ravvicinati
+        # e top_thresh*0.5 e' troppo aggressivo: scarta varianti valide che
+        # sarebbero state riprese al full-res. Stessa cosa per
+        # coarse_angle_factor (skip 1 ogni 2): con step fine non e' utile.
+        # Risultato osservato: precisione "veloce" 10° dava risultati
+        # migliori di "preciso" 2° proprio perche evitava il pruning.
+        eff_step = self._effective_angle_step()
+        top_factor = self.top_score_factor
+        cf_eff = max(1, coarse_angle_factor)
+        if eff_step <= 3.0:
+            top_factor = max(top_factor, 0.7)
+            cf_eff = 1
+        top_thresh = min_score * top_factor
 
         tw, th = self.template_size
         density_top = _jit_popcount(spread_top)
@@ -859,7 +1399,7 @@ class LineShapeMatcher:
 
         coarse_idx_list: list[int] = []  # varianti da valutare al top
         neighbor_map: dict[int, list[int]] = {}  # vi_coarse -> indici vicini
-        cf = max(1, coarse_angle_factor)
+        cf = cf_eff
         for scale_key, vi_list in variants_by_scale.items():
             vi_sorted = sorted(vi_list, key=lambda i: self.variants[i].angle_deg)
             n = len(vi_sorted)
@@ -980,8 +1520,8 @@ class LineShapeMatcher:
         # Full-res (parallelizzato) con bitmap
         spread0 = self._spread_bitmap(gray0)
         bit_active_full = int(
-            sum(1 << b for b in range(N_BINS)
-                if (spread0 & np.uint8(1 << b)).any())
+            sum(1 << b for b in range(self._n_bins)
+                if (spread0 & (spread0.dtype.type(1) << b)).any())
         )
         density_full = _jit_popcount(spread0)
         for sc in unique_scales:
@@ -1118,6 +1658,13 @@ class LineShapeMatcher:
                     search_radius=self._effective_angle_step() / 2.0,
                     original_score=score,
                 )
+            # Halcon SubPixel='least_squares_high': refinement iterativo
+            # gradient-field per precisione 0.05 px (vs 0.5 quadratic 2D).
+            if subpixel_lm and self.template_gray is not None:
+                cx_lm, cy_lm = self._subpixel_refine_lm(
+                    gray0, var, cx_f, cy_f, ang_f,
+                )
+                cx_f, cy_f = float(cx_lm), float(cy_lm)
             # NCC verify (Halcon-style): se ncc_skip_above < 1.0 salta
             # il calcolo per shape-score gia alti. Default 1.01 = NCC sempre,
             # piu sicuro contro falsi positivi (lo shape-score satura facile).
@@ -1126,10 +1673,37 @@ class LineShapeMatcher:
             # ranking/visualizzazione (uno score 1.0 vero richiede sia
             # match shape sia template gray identici).
             if verify_ncc and float(score_f) < ncc_skip_above:
-                ncc = self._verify_ncc(gray0, cx_f, cy_f, ang_f, var.scale)
+                ncc = self._verify_ncc(
+                    gray0, cx_f, cy_f, ang_f, var.scale,
+                    view_idx=getattr(var, "view_idx", 0),
+                )
                 if ncc < verify_threshold:
                     continue
                 score_f = (float(score_f) + max(0.0, ncc)) * 0.5
+            # Soft-margin gradient similarity: sostituisce o integra lo
+            # score con metric continua (cos sim gradients) invece di
+            # bin discreto. Halcon-style: piu robusto a piccole rotazioni.
+            if use_soft_score:
+                soft = self._compute_soft_score(
+                    gray0, var, cx_f, cy_f, ang_f,
+                )
+                score_f = (float(score_f) + soft) * 0.5
+            # Re-check min_score sullo score finale: NCC averaging puo
+            # abbattere lo shape-score sotto la soglia user. Senza questo
+            # check apparivano match con score < min_score (UI confusing).
+            if float(score_f) < min_score:
+                continue
+
+            # Feature recall (Halcon MinScore-style): conta quante feature
+            # template effettivamente combaciano nello spread scena alla
+            # pose finale. Scarta se sotto min_recall (default 0 = off).
+            # Util contro match parziali ad alto NCC ma poche feature reali.
+            if min_recall > 0.0:
+                recall = self._compute_recall(
+                    spread0, var, cx_f, cy_f, ang_f,
+                )
+                if recall < min_recall:
+                    continue
 
             # Ri-traslo coord da spazio crop ROI a spazio scena originale.
             cx_out = cx_f + roi_offset[0]
@@ -1137,17 +1711,42 @@ class LineShapeMatcher:
             poly = _oriented_bbox_polygon(
                 cx_out, cy_out, tw * var.scale, th * var.scale, ang_f,
             )
+            # Reject match con bbox che sfora pesantemente la scena:
+            # spesso indica match spurio (centro derivato male o scala
+            # incoerente). Tollera 25% out-of-bounds, sopra rigetta.
+            H_scn, W_scn = gray_full.shape
+            poly_area = float(cv2.contourArea(poly))
+            if poly_area > 0:
+                # Clip poly alla scena: intersezione con rettangolo (0,0,W,H)
+                scene_rect = np.array([
+                    [0, 0], [W_scn, 0], [W_scn, H_scn], [0, H_scn],
+                ], dtype=np.float32)
+                inter, _ = cv2.intersectConvexConvex(
+                    poly.astype(np.float32), scene_rect,
+                )
+                inside_ratio = float(inter) / poly_area
+                if inside_ratio < 0.75:
+                    continue
             # Penalità scala opzionale: score degrada con distanza da 1.0
             if scale_penalty > 0.0 and var.scale != 1.0:
                 score_f = float(score_f) * max(
                     0.0, 1.0 - scale_penalty * abs(var.scale - 1.0)
                 )
-            # NMS post-refine: refine puo spostare il match di nms_radius;
-            # ricontrollo overlap su match gia accettati per evitare
-            # duplicati (stesso oggetto trovato da varianti angolari diverse).
+            # NMS post-refine cross-variant: usa IoU bbox-poligonale invece
+            # di sola distanza centro. Due match orientati diversi ma vicini
+            # (pezzi adiacenti) NON vengono fusi se l'overlap reale e basso;
+            # due match dello stesso pezzo (centri uguali, rotazione simile)
+            # hanno IoU alto e vengono droppati.
+            # Fallback distanza centro per match con bbox degenere.
             dup = False
             for k in kept:
-                if (k.cx - cx_out) ** 2 + (k.cy - cy_out) ** 2 < r2:
+                iou = _poly_iou(k.bbox_poly, poly)
+                if iou > nms_iou_threshold:
+                    dup = True
+                    break
+                # Sicurezza: centri molto vicini (dentro nms_radius/2)
+                # sempre fusi, anche con orientamenti molto diversi.
+                if (k.cx - cx_out) ** 2 + (k.cy - cy_out) ** 2 < (r2 / 4.0):
                     dup = True
                     break
             if dup:

pm2d/web/server.py

@@ -48,6 +48,10 @@ IMAGES_DIR = Path(_images_dir_raw)
 if not IMAGES_DIR.is_absolute():
     IMAGES_DIR = PROJECT_ROOT / IMAGES_DIR
 
+# Cartella ricette pre-trained (V feature: save/load matcher)
+RECIPES_DIR = PROJECT_ROOT / "recipes"
+RECIPES_DIR.mkdir(exist_ok=True)
+
 from pm2d.line_matcher import LineShapeMatcher, Match
 from pm2d.auto_tune import auto_tune
 
@@ -249,9 +253,9 @@ PRECISION_ANGLE_STEP = {
 # Un operatore sceglie il livello di rigore, non un numero astratto.
 FILTRO_FP_MAP = {
     "off":      0.0,   # disabilitato: mantieni tutti i match shape-based
-    "leggero":  0.20,  # tollera variazioni intensità/illuminazione forti
-    "medio":    0.35,  # default bilanciato (consigliato)
-    "forte":    0.50,  # scarta match con intensità molto diversa dal template
+    "leggero":  0.30,  # tollera variazioni intensità/illuminazione forti
+    "medio":    0.50,  # default bilanciato (consigliato)
+    "forte":    0.70,  # scarta match con intensità molto diversa dal template
 }
 
 
@@ -267,6 +271,20 @@ class SimpleMatchParams(BaseModel):
     penalita_scala: float = 0.0       # 0 = score shape invariante, >0 = penalizza scala != 1
     min_score: float = 0.65
     max_matches: int = 25
+    # --- Halcon-mode flags (default off = backward compat) ---
+    # Init-time (richiede ri-train se cambiato)
+    use_polarity: bool = False        # F: 16 bin orientation mod 2pi
+    use_gpu: bool = False             # R: OpenCL UMat (silent fallback)
+    # Find-time (no retrain)
+    min_recall: float = 0.0           # M: filtra match con poche feature combaciate
+    use_soft_score: bool = False      # Y: cosine sim continua dei gradients
+    subpixel_lm: bool = False         # Z: precisione 0.05 px
+    nms_iou_threshold: float = 0.3    # A: IoU bbox poligonale
+    coarse_stride: int = 1            # sub-sampling top-level (>=1)
+    pyramid_propagate: bool = False   # propagazione candidati top->full
+    greediness: float = 0.0           # early-exit kernel (0..1)
+    refine_pose_joint: bool = False   # Nelder-Mead 3D (cx, cy, angle)
+    search_roi: list[int] | None = None  # [x, y, w, h] limita area
 
 
 def _simple_to_technical(
@@ -526,6 +544,9 @@ def match_simple(p: SimpleMatchParams):
     tech = _simple_to_technical(p, roi_img)
 
     key = _matcher_cache_key(roi_img, tech)
+    # Halcon-mode init params: incidono sul training, includere in cache key
+    halcon_init_key = f"|pol={p.use_polarity}|gpu={p.use_gpu}"
+    key = key + halcon_init_key
     m = _cache_get_matcher(key)
     if m is None:
         m = LineShapeMatcher(
@@ -537,17 +558,30 @@ def match_simple(p: SimpleMatchParams):
             scale_step=tech["scale_step"],
             spread_radius=tech["spread_radius"],
             pyramid_levels=tech["pyramid_levels"],
+            use_polarity=p.use_polarity,
+            use_gpu=p.use_gpu,
         )
         t0 = time.time(); n = m.train(roi_img); t_train = time.time() - t0
         _cache_put_matcher(key, m)
     else:
         n = len(m.variants); t_train = 0.0
     nms = tech["nms_radius"] if tech["nms_radius"] > 0 else None
+    search_roi_t = tuple(p.search_roi) if p.search_roi else None
     t0 = time.time()
     matches = m.find(
         scene, min_score=tech["min_score"], max_matches=tech["max_matches"],
         nms_radius=nms, verify_threshold=tech["verify_threshold"],
         scale_penalty=tech.get("scale_penalty", 0.0),
+        # Halcon-mode flags
+        min_recall=p.min_recall,
+        use_soft_score=p.use_soft_score,
+        subpixel_lm=p.subpixel_lm,
+        nms_iou_threshold=p.nms_iou_threshold,
+        coarse_stride=p.coarse_stride,
+        pyramid_propagate=p.pyramid_propagate,
+        greediness=p.greediness,
+        refine_pose_joint=p.refine_pose_joint,
+        search_roi=search_roi_t,
     )
     t_find = time.time() - t0
 
@@ -576,6 +610,70 @@ def tune(p: TuneParams):
     return {k: v for k, v in t.items() if not k.startswith("_")}
 
 
+# --- V: Save/Load ricette pre-trained ---
+
+class SaveRecipeParams(BaseModel):
+    model_id: str
+    scene_id: str | None = None
+    roi: list[int]
+    # Riusa stessi param simple per training equivalente
+    tipo: str = "intero"
+    simmetria: str = "nessuna"
+    scala: str = "fissa"
+    precisione: str = "normale"
+    use_polarity: bool = False
+    use_gpu: bool = False
+    name: str  # nome file ricetta (no path)
+
+
+@app.post("/recipes")
+def save_recipe(p: SaveRecipeParams):
+    """Allena matcher e salva su disco come ricetta riutilizzabile."""
+    model = _load_image(p.model_id)
+    if model is None:
+        raise HTTPException(404, "Modello non trovato")
+    x, y, w, h = p.roi
+    roi_img = model[y:y + h, x:x + w]
+    sp = SimpleMatchParams(
+        model_id=p.model_id, scene_id=p.scene_id or p.model_id, roi=p.roi,
+        tipo=p.tipo, simmetria=p.simmetria, scala=p.scala,
+        precisione=p.precisione,
+        use_polarity=p.use_polarity, use_gpu=p.use_gpu,
+    )
+    tech = _simple_to_technical(sp, roi_img)
+    m = LineShapeMatcher(
+        num_features=tech["num_features"],
+        weak_grad=tech["weak_grad"], strong_grad=tech["strong_grad"],
+        angle_range_deg=(tech["angle_min"], tech["angle_max"]),
+        angle_step_deg=tech["angle_step"],
+        scale_range=(tech["scale_min"], tech["scale_max"]),
+        scale_step=tech["scale_step"],
+        spread_radius=tech["spread_radius"],
+        pyramid_levels=tech["pyramid_levels"],
+        use_polarity=p.use_polarity,
+        use_gpu=p.use_gpu,
+    )
+    m.train(roi_img)
+    safe_name = "".join(c for c in p.name if c.isalnum() or c in "._-")
+    if not safe_name:
+        raise HTTPException(400, "Nome ricetta non valido")
+    if not safe_name.endswith(".npz"):
+        safe_name += ".npz"
+    target = RECIPES_DIR / safe_name
+    m.save_model(str(target))
+    return {"name": safe_name, "size": target.stat().st_size,
+            "n_variants": len(m.variants)}
+
+
+@app.get("/recipes")
+def list_recipes():
+    files = []
+    if RECIPES_DIR.is_dir():
+        for f in sorted(RECIPES_DIR.glob("*.npz")):
+            files.append({"name": f.name, "size": f.stat().st_size})
+    return {"files": files, "dir": str(RECIPES_DIR)}
+
+
 # Mount static
 app.mount("/static", StaticFiles(directory=STATIC_DIR), name="static")
 

pm2d/web/static/app.js

@@ -52,6 +52,39 @@ function readUserParams() {
       document.getElementById("p-penalita-scala").value),
     min_score: parseFloat(document.getElementById("p-min-score").value),
     max_matches: parseInt(document.getElementById("p-max-matches").value, 10),
+    ...readHalconFlags(),
+  };
+}
+
+function readHalconFlags() {
+  // Halcon-mode toggle: tutti i flag default-off, esposti via "Modalità Halcon"
+  const $cb = (id) => document.getElementById(id)?.checked ?? false;
+  const $num = (id, def) => {
+    const v = parseFloat(document.getElementById(id)?.value);
+    return Number.isFinite(v) ? v : def;
+  };
+  const $int = (id, def) => {
+    const v = parseInt(document.getElementById(id)?.value, 10);
+    return Number.isFinite(v) ? v : def;
+  };
+  const roiStr = document.getElementById("hc-search-roi")?.value.trim() ?? "";
+  let search_roi = null;
+  if (roiStr) {
+    const p = roiStr.split(/[ ,;]+/).map((x) => parseInt(x, 10));
+    if (p.length === 4 && p.every((v) => Number.isFinite(v))) search_roi = p;
+  }
+  return {
+    use_polarity: $cb("hc-use-polarity"),
+    use_gpu: $cb("hc-use-gpu"),
+    use_soft_score: $cb("hc-soft-score"),
+    subpixel_lm: $cb("hc-subpixel-lm"),
+    refine_pose_joint: $cb("hc-refine-joint"),
+    pyramid_propagate: $cb("hc-pyr-propagate"),
+    min_recall: $num("hc-min-recall", 0),
+    nms_iou_threshold: $num("hc-nms-iou", 0.3),
+    greediness: $num("hc-greediness", 0),
+    coarse_stride: $int("hc-coarse-stride", 1),
+    search_roi: search_roi,
   };
 }
 
@@ -294,12 +327,17 @@ async function doMatch() {
     const SCALE_MAP = {fissa:[1,1,0.1], mini:[0.9,1.1,0.05],
                        medio:[0.75,1.25,0.05], max:[0.5,1.5,0.05]};
     const PREC_MAP = {veloce:10, normale:5, preciso:2};
-    const FP_MAP = {off:0, leggero:0.20, medio:0.35, forte:0.50};
+    // Allineato a FILTRO_FP_MAP server-side (server.py)
+    const FP_MAP = {off:0, leggero:0.30, medio:0.50, forte:0.70};
     const [smin, smax, sstep] = SCALE_MAP[user.scala];
+    // NB: SYM_MAP[invariante]=0 e' valido (zero rotazioni). Uso ?? per
+    // distinguere "chiave mancante" da "valore zero": altrimenti 0 || 360
+    // collassa invariante a 360 = bug "simmetria non ha effetto".
+    const angMax = SYM_MAP[user.simmetria] ?? 360;
     body = {
       model_id: state.model.id, scene_id: state.scene.id, roi: state.roi,
-      angle_min: 0, angle_max: SYM_MAP[user.simmetria] || 360,
-      angle_step: PREC_MAP[user.precisione] || 5,
+      angle_min: 0, angle_max: angMax,
+      angle_step: PREC_MAP[user.precisione] ?? 5,
       scale_min: smin, scale_max: smax, scale_step: sstep,
       min_score: user.min_score, max_matches: user.max_matches,
       num_features: adv.num_features ?? 96,
@@ -307,7 +345,7 @@ async function doMatch() {
       strong_grad: adv.strong_grad ?? 60,
       spread_radius: adv.spread_radius ?? 5,
       pyramid_levels: adv.pyramid_levels ?? 3,
-      verify_threshold: adv.verify_threshold ?? (FP_MAP[user.filtro_fp] ?? 0.35),
+      verify_threshold: adv.verify_threshold ?? (FP_MAP[user.filtro_fp] ?? 0.50),
       nms_radius: adv.nms_radius ?? 0,
     };
   } else {
@@ -362,6 +400,44 @@ function setStatus(s) {
 }
 
 // ---------- Init ----------
+// ---------- V: Save recipe ----------
+async function saveRecipe() {
+  if (!state.model || !state.roi) {
+    alert("Seleziona modello e disegna ROI prima di salvare la ricetta.");
+    return;
+  }
+  const name = document.getElementById("hc-recipe-name").value.trim();
+  if (!name) {
+    alert("Inserisci un nome per la ricetta.");
+    return;
+  }
+  const user = readUserParams();
+  const body = {
+    model_id: state.model.id,
+    scene_id: state.scene?.id || state.model.id,
+    roi: state.roi,
+    tipo: user.tipo,
+    simmetria: user.simmetria,
+    scala: user.scala,
+    precisione: user.precisione,
+    use_polarity: user.use_polarity,
+    use_gpu: user.use_gpu,
+    name: name,
+  };
+  try {
+    const r = await fetch("/recipes", {
+      method: "POST",
+      headers: { "Content-Type": "application/json" },
+      body: JSON.stringify(body),
+    });
+    if (!r.ok) throw new Error(await r.text());
+    const j = await r.json();
+    alert(`Ricetta salvata: ${j.name}\n${j.n_variants} varianti, ${j.size} bytes`);
+  } catch (e) {
+    alert(`Errore salvataggio: ${e.message}`);
+  }
+}
+
 window.addEventListener("DOMContentLoaded", async () => {
   buildAdvancedForm();
   setupROI();
@@ -389,6 +465,8 @@ window.addEventListener("DOMContentLoaded", async () => {
     e.target.value = "";  // consente re-upload stesso file
   });
   document.getElementById("btn-match").addEventListener("click", doMatch);
+  document.getElementById("btn-save-recipe").addEventListener("click",
+    saveRecipe);
   const slider = document.getElementById("p-min-score");
   slider.addEventListener("input", (e) => {
     document.getElementById("v-score").textContent =

pm2d/web/static/index.html

@@ -129,6 +129,67 @@
       <input type="number" id="p-max-matches" value="25" min="1" max="200">
     </div>
 
+    <details>
+      <summary>Modalità Halcon</summary>
+      <div class="halcon-grid">
+        <label class="hc-row" title="16-bin orientation polarity-aware (mod 2π)">
+          <input type="checkbox" id="hc-use-polarity">
+          <span>Polarity 16-bin (F)</span>
+        </label>
+        <label class="hc-row" title="Score continuo cos(θ_t-θ_s) invece di bin">
+          <input type="checkbox" id="hc-soft-score">
+          <span>Soft-margin score (Y)</span>
+        </label>
+        <label class="hc-row" title="Sub-pixel refinement gradient field LM">
+          <input type="checkbox" id="hc-subpixel-lm">
+          <span>Sub-pixel LM 0.05 px (Z)</span>
+        </label>
+        <label class="hc-row" title="Refine congiunto Nelder-Mead (cx,cy,θ)">
+          <input type="checkbox" id="hc-refine-joint">
+          <span>Refine pose joint</span>
+        </label>
+        <label class="hc-row" title="Pyramid candidates propagation">
+          <input type="checkbox" id="hc-pyr-propagate">
+          <span>Pyramid propagate</span>
+        </label>
+        <label class="hc-row" title="OpenCL GPU offload (silent fallback CPU)">
+          <input type="checkbox" id="hc-use-gpu">
+          <span>GPU OpenCL (R)</span>
+        </label>
+
+        <div class="hc-row hc-num">
+          <label>Min recall (M)</label>
+          <input type="number" id="hc-min-recall" value="0.0" min="0" max="1" step="0.05">
+        </div>
+        <div class="hc-row hc-num">
+          <label>NMS IoU thr (A)</label>
+          <input type="number" id="hc-nms-iou" value="0.3" min="0" max="1" step="0.05">
+        </div>
+        <div class="hc-row hc-num">
+          <label>Greediness</label>
+          <input type="number" id="hc-greediness" value="0.0" min="0" max="1" step="0.1">
+        </div>
+        <div class="hc-row hc-num">
+          <label>Coarse stride</label>
+          <input type="number" id="hc-coarse-stride" value="1" min="1" max="4" step="1">
+        </div>
+        <div class="hc-row hc-num" style="grid-column:1/-1">
+          <label title="Limita area di ricerca scena: x,y,w,h (vuoto = tutta scena)">
+            Search ROI (x,y,w,h)
+          </label>
+          <input type="text" id="hc-search-roi" placeholder="es. 100,50,800,400">
+        </div>
+
+        <div class="hc-row" style="grid-column:1/-1; border-top:1px solid #444; padding-top:8px">
+          <label>Ricetta pre-trained (V)</label>
+          <div style="display:flex; gap:6px; margin-top:4px">
+            <input type="text" id="hc-recipe-name" placeholder="nome_ricetta" style="flex:1">
+            <button class="btn" id="btn-save-recipe" type="button">💾 Salva</button>
+          </div>
+        </div>
+      </div>
+    </details>
+
     <details>
       <summary>Avanzate</summary>
       <div id="adv-form"></div>

pm2d/web/static/style.css

@@ -156,3 +156,20 @@ footer h2 {
 }
 
 #col-model, #col-scene { min-width: 0; }
+
+/* Halcon-mode panel */
+.halcon-grid {
+  display: grid;
+  grid-template-columns: 1fr 1fr;
+  gap: 6px 12px;
+  margin-top: 6px;
+  font-size: 12px;
+}
+.hc-row {
+  display: flex; align-items: center; gap: 6px;
+}
+.hc-row.hc-num {
+  flex-direction: column; align-items: flex-start;
+}
+.hc-row.hc-num label { font-size: 11px; color: #aaa; }
+.hc-row.hc-num input { width: 100%; }