Spaces:
Running
Running
| import cv2 | |
| from cv2 import dnn_superres | |
| import numpy as np | |
| def enhance_photo_web(input_path, output_path, factor=4, face_restoration=True, color_correction=True): | |
| try: | |
| print(f"🧠 AI Engine starting: Reading {input_path}") | |
| img = cv2.imread(input_path) | |
| if img is None: | |
| return False | |
| # 1. THE RELAXED MEMORY SHIELD | |
| max_dim = 1600 | |
| height, width = img.shape[:2] | |
| if max(height, width) > max_dim: | |
| scale = max_dim / max(height, width) | |
| img = cv2.resize(img, (int(width * scale), int(height * scale)), interpolation=cv2.INTER_LANCZOS4) | |
| # 🌟 2. SAFE COLOR CORRECTION | |
| if color_correction: | |
| # We gently lift shadows without crushing the blacks or muddying the image | |
| lab = cv2.cvtColor(img, cv2.COLOR_BGR2LAB) | |
| l, a, b = cv2.split(lab) | |
| # Lowered the clipLimit to 1.2. This prevents the "dirty" look on selfies! | |
| clahe = cv2.createCLAHE(clipLimit=1.2, tileGridSize=(8,8)) | |
| cl = clahe.apply(l) | |
| limg = cv2.merge((cl,a,b)) | |
| img = cv2.cvtColor(limg, cv2.COLOR_LAB2BGR) | |
| # (Notice I completely removed the cv2.detailEnhance line that was ruining it!) | |
| # 3. FAST AI UPSCALING (FSRCNN) | |
| print("⚡ Booting up FSRCNN Neural Network...") | |
| sr = dnn_superres.DnnSuperResImpl_create() | |
| model_path = "FSRCNN_x4.pb" | |
| sr.readModel(model_path) | |
| sr.setModel("fsrcnn", 4) | |
| result = sr.upsample(img) | |
| # ✨ 4. CLEAN SHARPENING | |
| if face_restoration: | |
| print("✨ Applying clean unsharp mask...") | |
| # The aggressive matrix made noisy photos look terrible. | |
| # This Gaussian approach is much safer for skin and faces. | |
| gaussian_blur = cv2.GaussianBlur(result, (5, 5), 0) | |
| result = cv2.addWeighted(result, 1.5, gaussian_blur, -0.5, 0) | |
| cv2.imwrite(output_path, result) | |
| print(f"✅ SUCCESS: Photo enhanced cleanly!") | |
| return True | |
| except Exception as e: | |
| print(f"❌ PYTHON AI ERROR: {str(e)}") | |
| return False |