Fix Bug 1: Relax macro detection thresholds - adaptive blur threshold (p40), add blur_uniformity as 5th indicator, lower bypass to 0.55, add debug logging

agents/modality_detector.py

@@ -1,13 +1,15 @@
 """
-FORENSIQ — Capture Modality Detector v2
+FORENSIQ — Capture Modality Detector v3
 
 Classifies images BEFORE forensic analysis. Pure content-based detection
 that works even when Gradio strips metadata (format=None, no EXIF).
 
-Key fix: detection works entirely from image pixel analysis, not metadata.
-Metadata signals are bonus evidence only.
+Key fix v3: Relaxed macro detection thresholds. blur_frac uses adaptive 
+threshold (p40 instead of p25) for more accurate blur region estimation.
+Added blur_uniformity as 5th macro indicator. Debug logging to stderr.
 """
 
+import sys
 import numpy as np
 from PIL import Image
 from scipy.ndimage import gaussian_filter, sobel
@@ -33,9 +35,6 @@ def detect_modality(img: Image.Image) -> ModalityResult:
     rgb = np.array(img.convert("RGB")).astype(np.float64)
     
     # ═══ CRITICAL PRE-CHECK: Bayer CFA pattern ═══════════════════════
-    # Real camera images have Bayer demosaicing traces: σ_green < σ_red ≈ σ_blue
-    # If this is absent, the image CANNOT be from a real camera sensor.
-    # This blocks portrait mode suppression from protecting AI images.
     noise_std = {}
     for c, nm in enumerate(["red", "green", "blue"]):
         ch = rgb[:, :, c]
@@ -52,8 +51,7 @@ def detect_modality(img: Image.Image) -> ModalityResult:
     
     # ═══ CONTENT-BASED DETECTION (works without metadata) ═════════════
     
-    # ── Portrait mode detection ───────────────────────────────────────
-    # Core signal: bimodal sharpness distribution (sharp fg + uniform blur bg)
+    # ── Sharpness analysis (shared by portrait and macro) ─────────────
     lap = np.array([[0, 1, 0], [1, -4, 1], [0, 1, 0]], dtype=np.float64)
     laplacian = convolve2d(gray, lap, mode="same", boundary="symm")
     sharpness = gaussian_filter(np.abs(laplacian), sigma=max(10, min(h, w) // 80))
@@ -69,10 +67,14 @@ def detect_modality(img: Image.Image) -> ModalityResult:
     
     # Sharp region detection
     sharp_thresh = p75
-    blur_thresh = p25
     sharp_region = sharpness > sharp_thresh
-    blur_region = sharpness < blur_thresh
     sharp_frac = float(np.mean(sharp_region))
+    
+    # Blur region: use ADAPTIVE threshold (p40 of sharpness)
+    # This gives a more meaningful blur fraction than the fixed p25 definition
+    # which is always ~25% by construction.
+    blur_adaptive_thresh = float(np.percentile(sharpness, 40))
+    blur_region = sharpness < blur_adaptive_thresh
     blur_frac = float(np.mean(blur_region))
     
     # Blur uniformity (computational blur is very uniform)
@@ -90,13 +92,13 @@ def detect_modality(img: Image.Image) -> ModalityResult:
     
     indicators["p95_sharpness"] = round(p95, 2)
     indicators["bimodal_ratio"] = round(bimodal_ratio, 3)
+    indicators["blur_frac"] = round(blur_frac, 3)
     indicators["blur_uniformity"] = round(blur_uniformity, 3)
     indicators["transition_abruptness"] = round(transition, 2)
     indicators["has_detail"] = has_detail
     
-    # Portrait mode scoring — requires BOTH content signals AND real camera evidence
+    # ── Portrait mode detection ───────────────────────────────────────
     # CRITICAL: If no Bayer pattern, this CANNOT be a smartphone portrait photo.
-    # AI images that happen to have blurred backgrounds must NOT get portrait suppression.
     can_be_portrait = has_detail and has_bayer
     
     portrait_score = 0.0
@@ -114,19 +116,17 @@ def detect_modality(img: Image.Image) -> ModalityResult:
         indicators["portrait_detected"] = True
     
     # ── Macro/DSLR shallow DoF detection ─────────────────────────────
-    # Macro photos have: extreme sharpness ratio (center vs edge), very high
-    # bimodal ratio, large uniform blur region, and Bayer pattern present.
-    # Key difference from portrait mode: sharpness ratio is much higher (>5)
-    # because macro lenses produce more extreme DoF than phone portrait mode.
+    # Macro photos: extreme sharpness ratio (center vs edge), very high
+    # bimodal ratio, large blur region, uniform bokeh, and optionally Bayer.
     
     # Compute center-vs-edge sharpness ratio
-    ch, cw = gray.shape
-    center_region = sharpness[ch//4:3*ch//4, cw//4:3*cw//4]
+    ch_s, cw_s = gray.shape
+    center_region = sharpness[ch_s//4:3*ch_s//4, cw_s//4:3*cw_s//4]
     edge_region = np.concatenate([
-        sharpness[:ch//4, :].ravel(),
-        sharpness[3*ch//4:, :].ravel(),
-        sharpness[:, :cw//4].ravel(),
-        sharpness[:, 3*cw//4:].ravel(),
+        sharpness[:ch_s//4, :].ravel(),
+        sharpness[3*ch_s//4:, :].ravel(),
+        sharpness[:, :cw_s//4].ravel(),
+        sharpness[:, 3*cw_s//4:].ravel(),
     ])
     center_sharp = float(np.percentile(center_region, 90))
     edge_sharp = float(np.mean(edge_region))
@@ -138,20 +138,40 @@ def detect_modality(img: Image.Image) -> ModalityResult:
     
     indicators["sharpness_ratio"] = round(sharpness_ratio, 2)
     indicators["bg_color_std"] = round(bg_color_std, 2)
+    indicators["center_sharp_p90"] = round(center_sharp, 2)
+    indicators["edge_sharp_mean"] = round(edge_sharp, 2)
     
+    # ── Macro scoring (5 indicators) ─────────────────────────────────
     macro_score = 0.0
+    macro_components = []
+    
     if has_detail and sharpness_ratio > 3.0:
-        macro_score += 0.25  # Extreme center/edge sharpness difference
-    if has_detail and blur_frac > 0.35:
-        macro_score += 0.2   # Large blur region
+        macro_score += 0.25
+        macro_components.append(f"ratio={sharpness_ratio:.1f}")
+    
+    if has_detail and blur_frac > 0.30:
+        # Relaxed from 0.35 to 0.30 — adaptive p40 threshold gives ~40% for macro
+        macro_score += 0.15
+        macro_components.append(f"blur={blur_frac:.2f}")
+    
     if has_detail and bimodal_ratio > 1.5:
-        macro_score += 0.2   # Strong bimodal sharpness
+        macro_score += 0.20
+        macro_components.append(f"bimodal={bimodal_ratio:.2f}")
+    
     if has_detail and bg_color_std < 40:
-        macro_score += 0.15  # Uniform background color (bokeh'd)
+        macro_score += 0.15
+        macro_components.append(f"bg_std={bg_color_std:.1f}")
+    
+    # NEW: blur uniformity — bokeh backgrounds are very uniform in blur intensity
+    if has_detail and blur_uniformity > 0.6:
+        macro_score += 0.15
+        macro_components.append(f"blur_uni={blur_uniformity:.2f}")
     
-    # Macro requires Bayer OR high EXIF evidence (Unsplash strips some Bayer)
-    # Allow macro without Bayer if other signals are very strong
-    if macro_score >= 0.6:
+    indicators["macro_score"] = round(macro_score, 3)
+    indicators["macro_components"] = macro_components
+    
+    # Macro requires Bayer OR high signal strength (Unsplash strips Bayer traces)
+    if macro_score >= 0.55:
         scores["MACRO_DSLR"] = macro_score
         indicators["macro_detected"] = True
     elif macro_score >= 0.4 and has_bayer:
@@ -253,17 +273,26 @@ def detect_modality(img: Image.Image) -> ModalityResult:
         indicators["safety_override"] = "Low-detail image — suppression disabled"
     
     # SAFETY GUARD 2: No Bayer pattern = not from a real camera sensor.
-    # Exception: MACRO_DSLR with very strong signals can bypass this
+    # Exception: MACRO_DSLR with strong signals can bypass this
     # (Unsplash CDN processing can strip Bayer traces from real DSLR photos)
     if not has_bayer and modality in ("PORTRAIT_MODE", "SMARTPHONE", "MESSAGING"):
         modality = "UNKNOWN"
         conf = 0.2
         indicators["safety_override"] = f"No Bayer CFA pattern (margin={bayer_margin:.3f}) — not from a real camera sensor. All suppression disabled."
-    elif not has_bayer and modality == "MACRO_DSLR" and scores.get("MACRO_DSLR", 0) < 0.6:
+    elif not has_bayer and modality == "MACRO_DSLR" and scores.get("MACRO_DSLR", 0) < 0.55:
         # Weak macro signal without Bayer — don't trust it
         modality = "UNKNOWN"
         conf = 0.2
-        indicators["safety_override"] = f"Macro signal weak + no Bayer — suppression disabled"
+        indicators["safety_override"] = f"Macro signal weak ({scores.get('MACRO_DSLR', 0):.2f}) + no Bayer — suppression disabled"
+    
+    # ═══ DEBUG LOGGING ════════════════════════════════════════════════
+    print(f"[MODALITY] detected={modality} conf={conf:.2f} scores={scores}", file=sys.stderr)
+    print(f"[MODALITY] has_bayer={has_bayer} bayer_margin={bayer_margin:.4f}", file=sys.stderr)
+    print(f"[MODALITY] macro_score={macro_score:.3f} components={macro_components}", file=sys.stderr)
+    print(f"[MODALITY] sharpness_ratio={sharpness_ratio:.2f} bimodal={bimodal_ratio:.3f} blur_frac={blur_frac:.3f} blur_uni={blur_uniformity:.3f} bg_std={bg_color_std:.2f}", file=sys.stderr)
+    print(f"[MODALITY] p95={p95:.2f} has_detail={has_detail}", file=sys.stderr)
+    if indicators.get("safety_override"):
+        print(f"[MODALITY] SAFETY OVERRIDE: {indicators['safety_override']}", file=sys.stderr)
     
     # ═══ BUILD ADJUSTMENTS ════════════════════════════════════════════
     
@@ -301,7 +330,6 @@ def _get_modality_adjustments(modality: str) -> dict:
             "PRNU Uniformity": 0.2,
             "Demosaic Interpolation": 0.4,
             # Bimodal content (sharp subject + smooth bokeh) creates extreme kurtosis
-            # This is physics, not AI sharpening. Suppress the ceiling trigger.
             "DCT Kurtosis": 0.1,
             "Wavelet Kurtosis": 0.1,
             "Spectral Slope 1/f²": 0.4,