版本1
https://github.com/facebookresearch/segment-anything?tab=readme-ov-file#model-checkpoints
最新的是2
https://github.com/facebookresearch/sam2
环境 cuda11.8 配置
全图检测
mpu5pm 最好和最差模型 图像原图
mpu5pm 最好和最差模型 图像压缩一倍
mpu2pm 最好和最差模型 图像压缩一倍
mpu2pm 最好和最差模型 图像原图
import numpy as np
import cv2
from segment_anything import SamAutomaticMaskGenerator, sam_model_registry
import torch
import time
'''
环境
win11 coinda - base python12官方源码
git clone https://github.com/facebookresearch/segment-anything.git
cd segment-anything权重文件需手动下载
wget https://dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pthPyTorch 需要单独安装
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu118依赖库
pip install matplotlib pycocotools opencv-python onnx onnxruntime
# 代码格式工具
pip install flake8 isort black mypy
pip install segment_anything[all]'''# 1. 设置设备（GPU/CPU）
device = "cuda" if torch.cuda.is_available() else "cpu"
print(f"Using device: {device}")def get_model_info(size_key):"""根据输入返回模型信息"""return model_mapping.get(size_key, ("未知", "未知"))# 定义模型映射字典
model_ = {"Mode2_4g": ("vit_h", "sam_vit_h_4b8939.pth"),"Mode1_2g": ("vit_l", "sam_vit_l_0b3195.pth"),"Mode300mb": ("vit_b", "sam_vit_b_01ec64.pth")
}# 1 图像缩放倍数
scalse=2
# 2 模型选择 最大2_4g 中间1_2g 最小300mb
mode_test='Mode300mb'
model_type, checkpoint_path = model_.get(mode_test, ("vit_b", "sam_vit_b_01ec64.pth")) # 默认数据# 1. 初始化模型
'''
vit_b
vit_l
vit_h
'''
#model_type = "vit_b" # 对应 sam_vit_b_01ec64.pth
'''
sam_vit_b_01ec64 300mb
sam_vit_l_0b3195 1.2g
sam_vit_h_4b8939 2.4g'''#checkpoint_path = "./sam_vit_b_01ec64.pth" # 假设权重文件在当前目录
sam = sam_model_registry[model_type](checkpoint=checkpoint_path)
#sam = sam_model_registry[model_type](checkpoint=checkpoint_path).to('cpu')
sam.to(device) # 将模型移至 GPU（如果可用）# 2. 配置自动分割生成器（参数可根据需求调整）
# mask_generator = SamAutomaticMaskGenerator(
# sam,
# points_per_side=40, # 高空图像需要更密集的点
# pred_iou_thresh=0.9, # 提高质量要求
# min_mask_region_area=200 # 过滤小区域
# )# #Q2: 分割结果过于碎片化
# # 调整参数
# mask_generator = SamAutomaticMaskGenerator(
# sam,
# points_per_side=20, # 减少点数
# pred_iou_thresh=0.9, # 提高质量阈值
# stability_score_thresh=0.95
# )mask_generator = SamAutomaticMaskGenerator(model=sam,points_per_side=32, # 每边生成的点数pred_iou_thresh=0.86, # 掩膜质量阈值stability_score_thresh=0.92, # 稳定性阈值crop_n_layers=0, # 不使用多层级裁剪crop_overlap_ratio=0.3, # 裁剪重叠比例min_mask_region_area=100 # 最小掩膜区域面积（像素）
)# 3. 加载并预处理图像
image_path = "npu5pm.JPG" # 测试图像路径
image = cv2.imread(image_path)height, width = image.shape[:2]
new_width = int(width // scalse)
new_height = int(height // scalse)image = cv2.resize(image, (new_width, new_height))
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB) # 转为RGB格式# 4. 生成分割掩膜
start_time = time.time()
masks = mask_generator.generate(image)
end_time = time.time()
print(f"预测耗时：{end_time - start_time:.4f} 秒") # 保留4位小数# 5. 可视化结果（可选）
output_image = image.copy()
for mask in masks:# 为每个掩膜随机生成颜色color = np.random.randint(0, 256, 3)# 半透明叠加掩膜区域output_image[mask['segmentation']] = color * 0.6 + output_image[mask['segmentation']] * 0.4#Q3: 需要保存掩膜为单独文件# for i, mask in enumerate(masks):
# cv2.imwrite(f"mask_{i}.png", mask['segmentation'].astype('uint8') * 255)#提取特定区域的原图内容# 提取面积最大的前3个对象
# masks_sorted = sorted(masks, key=lambda x: x['area'], reverse=True)[:3]# for i, mask in enumerate(masks_sorted):
# # 创建透明背景的PNG
# rgba = cv2.cvtColor(image, cv2.COLOR_RGB2RGBA)
# rgba[~mask['segmentation']] = [0,0,0,0] # 非掩膜区域透明
# cv2.imwrite(f"object_{i}.png", rgba)'''
输出结果解析：
每个 mask包含：
segmentation: 二值掩膜矩阵
area: 区域像素面积
bbox: [x,y,width,height] 边界框
predicted_iou: 模型预测的质量分数
stability_score: 稳定性评分
'''# 保存结果
cv2.imwrite("Result_"+mode_test+"_"+str(scalse)+"_"+image_path, cv2.cvtColor(output_image, cv2.COLOR_RGB2BGR))# 打印分割结果统计
print(f"检测到 {len(masks)} 个分割区域")
print("第一个掩膜信息示例:", {k: v for k, v in masks[0].items() if k != 'segmentation'})