Download removaEnhancer.py from andevs/bgremove: direct link, hf CLI and curl.
- Browser
- Download file 2.91 kB
-
https://huggingface.co/spaces/andevs/bgremove/resolve/main/removaEnhancer.py
- Command line
-
hf download hf://spaces/andevs/bgremove/removaEnhancer.py
-
curl -L -o removaEnhancer.py https://huggingface.co/spaces/andevs/bgremove/resolve/main/removaEnhancer.py
2.91 kB
| """CPU-conscious, conservative alpha enhancement for Remova images and video. | |
| Does not hallucinate missing objects; preserves high-confidence foreground and | |
| avoids processing every frame with a second neural network. | |
| """ | |
| import cv2 | |
| import numpy as np | |
| from PIL import Image | |
| def enhance_rgba(result: Image.Image, source: Image.Image, model: str = 'silueta', | |
| video: bool = False) -> Image.Image: | |
| """Edge-aware refinement restricted to uncertain alpha pixels.""" | |
| rgba = np.asarray(result.convert('RGBA')).copy() | |
| alpha = rgba[:, :, 3].copy() | |
| if alpha.size < 64 or alpha.max() < 8 or alpha.min() > 247: | |
| return result | |
| guide = np.asarray(source.convert('RGB')) | |
| if guide.shape[:2] != alpha.shape: | |
| guide = cv2.resize(guide, (alpha.shape[1], alpha.shape[0]), interpolation=cv2.INTER_AREA) | |
| # A one-channel guided filter avoids colour-channel covariance artifacts. | |
| gray = cv2.cvtColor(guide, cv2.COLOR_RGB2GRAY).astype(np.float32) / 255.0 | |
| p = alpha.astype(np.float32) / 255.0 | |
| radius = 3 if video else 4 | |
| kernel = (radius * 2 + 1, radius * 2 + 1) | |
| mean_i = cv2.boxFilter(gray, -1, kernel, normalize=True) | |
| mean_p = cv2.boxFilter(p, -1, kernel, normalize=True) | |
| cov = cv2.boxFilter(gray * p, -1, kernel, normalize=True) - mean_i * mean_p | |
| var = cv2.boxFilter(gray * gray, -1, kernel, normalize=True) - mean_i * mean_i | |
| a = cov / (var + 0.002) | |
| b = mean_p - a * mean_i | |
| refined = cv2.boxFilter(a, -1, kernel) * gray + cv2.boxFilter(b, -1, kernel) | |
| refined = np.clip(refined, 0, 1) | |
| # Only adjust the uncertain band. Do not erode hair or opaque body regions. | |
| uncertain = (alpha > 12) & (alpha < 243) | |
| strength = 0.42 if model == 'silueta' else 0.32 | |
| if video: | |
| strength *= 0.8 | |
| output = p.copy() | |
| output[uncertain] = (1 - strength) * p[uncertain] + strength * refined[uncertain] | |
| # Tiny isolated hard-mask speckles can be removed only when sufficiently | |
| # far from any substantial component. Keep multi-object subjects intact. | |
| if not video: | |
| core = (alpha >= 200).astype(np.uint8) | |
| count, labels, stats, _ = cv2.connectedComponentsWithStats(core, 8) | |
| if count > 2: | |
| large = np.zeros_like(core) | |
| threshold = max(16, int(alpha.size * 0.00008)) | |
| for i in range(1, count): | |
| if stats[i, cv2.CC_STAT_AREA] >= threshold: | |
| large[labels == i] = 1 | |
| if large.any(): | |
| near = cv2.dilate(large, np.ones((7, 7), np.uint8)) | |
| for i in range(1, count): | |
| if stats[i, cv2.CC_STAT_AREA] < threshold: | |
| component = labels == i | |
| if not np.any(near & component): | |
| output[component] = 0 | |
| rgba[:, :, 3] = np.uint8(np.clip(output * 255 + 0.5, 0, 255)) | |
| return Image.fromarray(rgba, 'RGBA') | |