simple_symbol_det - 符号检测模型 (昇腾 NPU 适配)

# 安装基础依赖
pip install -i https://pypi.tuna.tsinghua.edu.cn/simple numpy opencv-python onnxruntime pyclipper shapely

# NPU 推理需要安装 onnxruntime-cann
# 请根据 CANN 版本选择对应的 onnxruntime-cann 版本
pip install onnxruntime-cann

cd /opt/atomgit/workspace/simple_symbol_det
python3 inference.py \
  --model_path /path/to/last.onnx \
  --image_path ./test_text.jpg \
  --provider CPUExecutionProvider \
  --draw ./result_cpu.jpg

python3 inference.py \
  --model_path /path/to/last.onnx \
  --image_path ./test_text.jpg \
  --provider CANNExecutionProvider \
  --draw ./result_npu.jpg

python3 compare_cpu_npu.py \
  --model_path /path/to/last.onnx \
  --image_path ./test_text.jpg \
  --output_dir ./results

import onnxruntime

# CPU 推理
cpu_session = onnxruntime.InferenceSession(
    model_path,
    providers=["CPUExecutionProvider"]
)

# NPU 推理
npu_session = onnxruntime.InferenceSession(
    model_path,
    providers=["CANNExecutionProvider"]
)

def preprocess(img, resized_shape=(768, 768)):
    ori_h, ori_w = img.shape[:2]
    target_h, target_w = resized_shape
    
    # 保持比例缩放并填充
    scale = min(target_h / ori_h, target_w / ori_w)
    new_size = (int(ori_w * scale), int(ori_h * scale))
    img_resized = cv2.resize(img, new_size)
    
    mask = np.ones((target_h, target_w), dtype=np.uint8) * 255
    mask[:img_resized.shape[0], :img_resized.shape[1]] = img_resized
    
    # 归一化并调整维度
    mask = mask.astype(np.float32) / 255.0
    mask = np.expand_dims(mask, 0)  # (1, H, W)
    mask = np.expand_dims(mask, 0)  # (1, 1, H, W)
    
    return mask, compress_ratios

def detect(model, image, thresh=0.2):
    # 推理
    batch, compress_ratios = preprocess(image)
    ort_outs = model.run(None, {model.get_inputs()[0].name: batch})
    pred = np.squeeze(ort_outs[0])
    
    # 二值化
    segmentation = ((pred > thresh) * 255).astype(np.uint8)
    
    # 轮廓检测
    contours, _ = cv2.findContours(segmentation, cv2.RETR_LIST, cv2.CHAIN_APPROX_SIMPLE)
    
    # 提取最小外接矩形
    boxes, scores = boxes_from_bitmap(pred, segmentation)
    
    # 坐标映射回原图
    for box in boxes:
        box[:, 0] *= image.shape[1] / resized_shape[1] / compress_ratios[1]
        box[:, 1] *= image.shape[0] / resized_shape[0] / compress_ratios[0]
    
    return boxes, scores

# NPU 推理
python3 inference.py --device npu

# CPU 推理
python3 inference.py --device cpu