import os import numpy as np from PIL import Image, ImageFilter, ImageChops def split_cloud_smart(): input_path = os.path.join(os.path.dirname(__file__), 'origin', 'cloud.jpg') output_dir = os.path.join(os.path.dirname(__file__), 'clouds') if not os.path.exists(output_dir): os.makedirs(output_dir) print(f"Processing {input_path}...") try: img = Image.open(input_path).convert("RGBA") except Exception as e: print(f"Error opening image: {e}") return # 1. 智能采样背景色 (Smart Sampling) # 取图像四周边缘的像素来计算背景特征,比只取一个角更稳健 width, height = img.size # 提取上、下、左、右四条边的像素 top_edge = np.array(img.crop((0, 0, width, 1))) bottom_edge = np.array(img.crop((0, height-1, width, height))) left_edge = np.array(img.crop((0, 0, 1, height))) right_edge = np.array(img.crop((width-1, 0, width, height))) # 合并边缘像素 edges = np.concatenate([ top_edge.reshape(-1, 4), bottom_edge.reshape(-1, 4), left_edge.reshape(-1, 4), right_edge.reshape(-1, 4) ]) # 计算背景色的平均值和标准差,用于确定容差范围 bg_mean = np.mean(edges, axis=0) print(f"Smart sampled background color (RGBA): {bg_mean}") # 2. HSV 色彩空间分离 (HSV Separation) # 将图片转为 HSV,利用饱和度(S)和亮度(V)来区分云(通常S低V高)和深色背景(通常S高V低) hsv_img = img.convert("HSV") hsv_data = np.array(hsv_img) rgb_data = np.array(img) # 提取通道 H, S, V = hsv_data[:,:,0], hsv_data[:,:,1], hsv_data[:,:,2] R, G, B = rgb_data[:,:,0], rgb_data[:,:,1], rgb_data[:,:,2] # 计算 RGB 欧氏距离 (针对平均背景色) # 只比较 RGB 前三个通道 diff_r = R.astype(float) - bg_mean[0] diff_g = G.astype(float) - bg_mean[1] diff_b = B.astype(float) - bg_mean[2] rgb_distance = np.sqrt(diff_r**2 + diff_g**2 + diff_b**2) # 定义阈值 # RGB 容差:允许背景有一定的颜色波动 rgb_tolerance = 60.0 # HSV 辅助判断: # 背景通常是深紫色:需要保护云朵(白色/灰色),云朵的特征是低饱和度(Low S) # 如果一个像素离背景色有点远,但它饱和度很高且偏紫,那它可能还是背景(渐变区) # 如果一个像素离背景色很近,但它饱和度极低(它是灰色的云边缘),那应该保留 # 创建 Alpha Mask (0 为完全透明/背景,255 为完全不透明/云) # 初始 Mask:距离背景色越近,Alpha 越小 alpha_mask = np.zeros_like(H, dtype=np.float32) # 核心逻辑: # 1. 主要是背景:RGB 距离 < 容差 # 2. 渐变增强:对于边缘区域,使用 Sigmoid 函数做软过渡,而不是硬切 # 归一化距离,距离越小越接近背景 normalized_dist = np.clip(rgb_distance / rgb_tolerance, 0, 1) # 简单的线性映射翻转:距离越小(背景),Alpha越小(透明) # 使用平滑函数 (Smoothstep) 让过渡更自然: 3x^2 - 2x^3 alpha_mask = np.clip((normalized_dist - 0.2) / 0.6, 0, 1) # 0.2到0.8之间过渡 alpha_mask = alpha_mask * alpha_mask * (3 - 2 * alpha_mask) # 3. 保护云朵核心 (Cloud Core Protection) # 如果像素很亮(V高)且饱和度很低(S低),强制认为是云,设为不透明 # 假设云是白色的,背景是深色的 is_cloud_core = (V > 150) & (S < 60) alpha_mask[is_cloud_core] = 1.0 # 4. 转换回 0-255 并应用羽化 final_alpha = (alpha_mask * 255).astype(np.uint8) # 创建蒙版图像 mask_img = Image.fromarray(final_alpha, mode='L') # 边缘羽化 (Matte Refinement) # 对蒙版进行轻微模糊,消除锯齿 mask_img = mask_img.filter(ImageFilter.GaussianBlur(radius=1.5)) # 将处理好的 Alpha 通道应用回原图 r, g, b, a = img.split() img_transparent = Image.merge('RGBA', (r, g, b, mask_img)) # 切割逻辑保持不变 width, height = img.size cell_width = width // 3 cell_height = height // 3 count = 0 for r in range(3): for c in range(3): left = c * cell_width top = r * cell_height right = left + cell_width bottom = top + cell_height cell = img_transparent.crop((left, top, right, bottom)) output_filename = f"cloud_{count}.png" output_path = os.path.join(output_dir, output_filename) cell.save(output_path) print(f"Saved {output_path}") count += 1 print("Smart split done!") if __name__ == "__main__": split_cloud_smart()