run.py

import os
import sys
import cv2
import torch
import numpy as np
from PIL import Image

# ==========================================
# 1. 环境配置
# ==========================================
try:
    from transformers import pipeline
except ImportError:
    print("错误: 缺少 transformers 库。")
    print("请运行: pip install transformers accelerate")
    sys.exit(1)

class Config:
    # 模型路径 (包含 config.json 和 safetensors 的文件夹)
    model_path = "/data/test/four_corn/sam3" 
    
    # 输入图片路径
    input_path = "/data/test/four_corn/Pakistan_card_img/ss_20260127153633_47_2.jpg"
    output_dir = "output_sam3_auto"
    
    # 设备配置
    device = "cuda" if torch.cuda.is_available() else "cpu"
    
    # 推理参数
    points_per_batch = 64 


# ==========================================
# 2. 筛选逻辑 (核心修复部分)
# ==========================================
def filter_masks(masks, img_area):
    """
    从 Mask 列表中找到最像“证件”的那个
    """
    candidates = []
    
    for mask_data in masks:
        raw_seg = mask_data['segmentation']
        
        # --- 格式统一化处理 (修复 TypeError 的关键) ---
        # 1. 如果是 PyTorch Tensor (常见于 CUDA 推理)，需转到 CPU 并转为 Numpy
        if isinstance(raw_seg, torch.Tensor):
            seg = raw_seg.detach().cpu().numpy()
        # 2. 如果是 PIL Image，转 Numpy
        elif isinstance(raw_seg, Image.Image):
            seg = np.array(raw_seg)
        # 3. 如果已经是 Numpy，直接用
        elif isinstance(raw_seg, np.ndarray):
            seg = raw_seg
        else:
            # 兜底：尝试强制转换
            seg = np.array(raw_seg)

        # 确保是 boolean 类型 (True/False)，因为有些输出是 0/255 或 0/1 的 int
        if seg.dtype != bool:
            seg = (seg > 0)

        # --- 以下逻辑保持不变 ---

        # 计算面积 (True 的像素总数)
        area = np.sum(seg)
        
        # 1. 面积初步筛选 (比如小于 5% 或大于 90% 过滤掉)
        if area < img_area * 0.05 or area > img_area * 0.90:
            continue
            
        # 2. 转为 uint8 以供 OpenCV 处理
        mask_uint8 = (seg * 255).astype(np.uint8)
        
        # 3. 计算轮廓和外接矩形
        contours, _ = cv2.findContours(mask_uint8, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
        if not contours: continue
        cnt = max(contours, key=cv2.contourArea)
        
        # 最小外接矩形 (Rotated Rect)
        rect = cv2.minAreaRect(cnt) 
        box_area = rect[1][0] * rect[1][1]
        
        if box_area == 0: continue

        # 4. 计算“矩形度” (Mask面积 / 外接矩形面积)
        rectangularity = area / box_area
        
        candidates.append({
            'mask': seg,                # 保存 boolean mask 用于绘图
            'rect': rect,               # ((cx, cy), (w, h), angle)
            'score': rectangularity * np.log(area), # 综合评分
            'box_points': cv2.boxPoints(rect)
        })
    
    # 按评分排序，取第一名
    if candidates:
        candidates.sort(key=lambda x: x['score'], reverse=True)
        return candidates[0]
    return None

# ==========================================
# 3. 主程序
# ==========================================
def main():
    cfg = Config()
    
    # 1. 加载模型
    print(f">>> 正在加载本地 SAM 3 模型: {cfg.model_path} ...")
    if not os.path.exists(cfg.model_path):
        print(f"错误: 路径不存在 {cfg.model_path}")
        return

    try:
        # 加载 mask-generation 管道
        generator = pipeline(
            "mask-generation",
            model=cfg.model_path,
            device=cfg.device,
            points_per_batch=cfg.points_per_batch
        )
    except Exception as e:
        print(f"模型加载失败: {e}")
        return

    print(" -> 模型加载完成")

    # 2. 读取图片
    img_cv2 = cv2.imread(cfg.input_path)
    if img_cv2 is None:
        print(f"图片读取失败: {cfg.input_path}")
        return
    
    img_h, img_w = img_cv2.shape[:2]
    img_area = img_h * img_w
    
    # 转为 PIL Image 供模型使用
    img_pil = Image.fromarray(cv2.cvtColor(img_cv2, cv2.COLOR_BGR2RGB))

    # 3. 自动分割全图
    print(f">>> 正在全图分割: {os.path.basename(cfg.input_path)} ...")
    
    # --- 运行推理 ---
    outputs = generator(img_pil)
    
    # --- 数据格式标准化 ---
    formatted_masks = []
    
    # 情况 A: outputs 是字典 (标准格式: {'masks': [...], 'scores': [...]})
    if isinstance(outputs, dict) and 'masks' in outputs:
        raw_masks = outputs['masks']
        for mask in raw_masks:
            formatted_masks.append({
                'segmentation': mask,
                'area': 0             # 占位
            })
            
    # 情况 B: outputs 是列表 (兼容旧格式)
    elif isinstance(outputs, list):
        for item in outputs:
            if isinstance(item, dict) and 'mask' in item:
                formatted_masks.append({'segmentation': item['mask'], 'area': 0})
            elif isinstance(item, Image.Image):
                formatted_masks.append({'segmentation': item, 'area': 0})
            elif isinstance(item, torch.Tensor): # 处理直接返回 Tensor 列表的情况
                formatted_masks.append({'segmentation': item, 'area': 0})
    
    print(f" -> 生成了 {len(formatted_masks)} 个掩膜片段")

    # 4. 筛选出证件
    best_candidate = filter_masks(formatted_masks, img_area)

    # 5. 绘图保存
    vis_img = img_cv2.copy()
    
    if best_candidate:
        print(">>> 找到最佳证件区域！")
        mask = best_candidate['mask']
        box_points = np.int64(best_candidate['box_points'])
        
        # 绿色 Mask
        vis_img[mask] = vis_img[mask] * 0.5 + np.array([0, 255, 0]) * 0.5
        
        # 红色 旋转矩形框
        cv2.drawContours(vis_img, [box_points], 0, (0, 0, 255), 3)
        
        # 角度
        angle = best_candidate['rect'][-1]
        print(f" -> 旋转角度: {angle:.2f}")
    else:
        print(">>> 未找到符合条件的证件区域")

    os.makedirs(cfg.output_dir, exist_ok=True)
    save_path = os.path.join(cfg.output_dir, "auto_sam3_" + os.path.basename(cfg.input_path))
    cv2.imwrite(save_path, vis_img)
    print(f"结果已保存: {save_path}")

if __name__ == "__main__":
    main()