zjsh_yolov11/angle_base_seg/bushu.py

import os
import cv2
import numpy as np
from rknnlite.api import RKNNLite

# ------------------- 参数 -------------------
objectThresh = 0.7  # 置信度阈值，可调整

# ------------------- 工具函数 -------------------
def letterbox_resize(image, size, bg_color=114):
    target_width, target_height = size
    h, w, _ = image.shape
    scale = min(target_width / w, target_height / h)
    new_w, new_h = int(w * scale), int(h * scale)
    resized = cv2.resize(image, (new_w, new_h), interpolation=cv2.INTER_AREA)
    canvas = np.ones((target_height, target_width, 3), dtype=np.uint8) * bg_color
    offset_x, offset_y = (target_width - new_w) // 2, (target_height - new_h) // 2
    canvas[offset_y:offset_y+new_h, offset_x:offset_x+new_w] = resized
    return canvas, scale, offset_x, offset_y

def sigmoid(x):
    return 1 / (1 + np.exp(-x))

# ------------------- 单目标分割函数 -------------------
def detect_single_mask(model_path, image_path):
    img = cv2.imread(image_path)
    if img is None:
        print(f"❌ 错误：无法读取图像 {image_path}")
        return None

    img_resized, scale, offset_x, offset_y = letterbox_resize(img, (640, 640))
    infer_img = img_resized[..., ::-1]  # BGR->RGB
    infer_img = np.expand_dims(infer_img, 0)

    # ------------------- RKNN 推理 -------------------
    rknn = RKNNLite(verbose=True)
    print('--> Load RKNN model')
    rknn.load_rknn(model_path)
    print('done')

    print('--> Init runtime environment')
    rknn.init_runtime(core_mask=RKNNLite.NPU_CORE_0)
    print('done')

    print('--> Running model')
    outputs = rknn.inference([infer_img])
    rknn.release()

    # ------------------- 处理输出 -------------------
    # outputs顺序参考你提供：
    # [output1, output2, output3, output4, ... , output12, output13]
    # 我们只取置信度最高点的 mask

    mask_coeffs_list = [outputs[3], outputs[7], outputs[11]]  # mask coefficients
    conf_list = [outputs[1], outputs[5], outputs[9]]            # object confidence
    proto = outputs[12][0]                                      # proto [32,160,160]

    # 找到所有尺度的最大置信度位置
    best_idx = None
    best_conf = -1
    best_scale = None
    for i, conf_map in enumerate(conf_list):
        conf_map_flat = conf_map.flatten()
        idx = np.argmax(conf_map_flat)
        if conf_map_flat[idx] > best_conf:
            best_conf = conf_map_flat[idx]
            best_idx = idx
            best_scale = i

    if best_conf < objectThresh:
        print(f"⚠️ 置信度低于阈值 {objectThresh}，未检测到目标")
        return None

    # ------------------- 构建 mask -------------------
    coeff = mask_coeffs_list[best_scale].reshape(mask_coeffs_list[best_scale].shape[1], -1)
    mask_flat = np.matmul(coeff[:, best_idx], proto.reshape(proto.shape[0], -1))
    mask = sigmoid(mask_flat).reshape(proto.shape[1], proto.shape[2])

    # resize 回原图
    mask_resized = cv2.resize(mask, (img.shape[1], img.shape[0]))
    mask_bin = (mask_resized > 0.5).astype(np.uint8) * 255

    # 保存或显示
    cv2.imwrite("mask_result.png", mask_bin)
    print("✅ 单目标 mask 已保存: mask_result.png")

    return mask_bin

# ------------------- 调用示例 -------------------
if __name__ == "__main__":
    model_path = "seg.rknn"
    image_path = "2.jpg"
    detect_single_mask(model_path, image_path)