feat: generate seamless RGBA overlay with transparent authentic background

api/model_server.py

@@ -273,7 +273,10 @@ class TamperingModelServer:
         # Directly evaluated from multi-frequency feature map, without artificial color or class overrides.
         mask_np = torch.sigmoid(mask)[0, 0].cpu().numpy()
 
-        # Convert raw continuous tampering probabilities [0.0, 1.0] to Turbo false-color RGB heatmap
+        peak_score = round(float(mask_np.max()), 4)
+        coverage_ratio = round(float(np.mean(mask_np > 0.50)), 4)
+
+        # 1. Standalone publication-style Turbo false-color RGB heatmap (Matches Figure 2 Column d in paper)
         # 0.0 (Deep Navy/Indigo): Pristine Authentic | 0.50 (Lime Green): Transition | 1.0 (Crimson Red): High Tampering
         lut_idx = np.clip(mask_np * 255.0, 0, 255).astype(np.int32)
         heatmap_rgb = TURBO_LUT[lut_idx]
@@ -282,6 +285,18 @@ class TamperingModelServer:
         mask_pil.save(buf, format="PNG")
         mask_b64 = base64.b64encode(buf.getvalue()).decode("utf-8")
 
+        # 2. Seamless RGBA Overlay: Authentic background is 100% transparent (alpha=0),
+        # only tampering hotspots glow in vivid Turbo colors without discoloring the original photograph
+        overlay_rgba = np.zeros((self.img_size, self.img_size, 4), dtype=np.uint8)
+        overlay_rgba[:, :, :3] = heatmap_rgb
+        # Threshold: authentic pixels (p < 0.12) are completely transparent; ramps smoothly up to 210 alpha for hotspots
+        alpha_map = np.clip((mask_np - 0.12) / 0.25 * 210.0, 0, 210).astype(np.uint8)
+        overlay_rgba[:, :, 3] = alpha_map
+        overlay_pil = Image.fromarray(overlay_rgba, mode="RGBA")
+        buf_overlay = io.BytesIO()
+        overlay_pil.save(buf_overlay, format="PNG")
+        overlay_b64 = base64.b64encode(buf_overlay.getvalue()).decode("utf-8")
+
         # Frequency bands
         bands = self.model.decomposer(tensor)
         band_mags = [float(b.abs().mean().item()) for b in bands]
@@ -310,6 +325,9 @@ class TamperingModelServer:
             "tampering_detected": pred_idx in [1, 2],
             "class_probabilities": class_probs,
             "tampering_heatmap_b64": mask_b64,
+            "tampering_overlay_b64": overlay_b64,
+            "tampering_peak_score": peak_score,
+            "tampering_coverage_ratio": coverage_ratio,
             "frequency_band_contributions": band_dict,
             "processing_time_ms": round(elapsed_ms, 2),
             "model_variant": self.variant,