zjsh_yolov11/yemian/yemian_line/seg_bushu/danmu_bushu.py

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

# ======================
# 配置
# ======================
IMAGE_PATH = "3.png"
MODEL_PATH = "segr.rknn"

OUT_OVERLAY = "result_overlay.jpg"
DEBUG_INPUT = "debug_input_roi.png"
DEBUG_PROTO = "debug_proto_mask.png"
DEBUG_INST_PROTO = "debug_inst_proto.png"

IMG_SIZE = 640
OBJ_THRESH = 0.25
MASK_THRESH = 0.5
STRIDES = [8, 16, 32]

ROIS = [
    (670, 623, 465, 178),
]

# ======================
# 工具函数
# ======================
def sigmoid(x):
    return 1 / (1 + np.exp(-x))


def resize_to_640(img):
    """seg 专用：禁止 letterbox，直接 resize"""
    return cv2.resize(img, (IMG_SIZE, IMG_SIZE), interpolation=cv2.INTER_LINEAR)


def dfl_decode(dfl):
    bins = np.arange(16)
    dfl = sigmoid(dfl)
    dfl /= np.sum(dfl, axis=1, keepdims=True)
    return np.sum(dfl * bins, axis=1)


def largest_cc(mask):
    num, labels = cv2.connectedComponents(mask.astype(np.uint8))
    if num <= 1:
        return mask
    areas = [(labels == i).sum() for i in range(1, num)]
    return (labels == (np.argmax(areas) + 1)).astype(np.uint8)


# ======================
# 单 ROI 推理（完整语义 mask）
# ======================
def infer_single_roi(rknn, roi):
    h0, w0 = roi.shape[:2]

    # ---------- 1️⃣ 正确的 seg 输入 ----------
    inp_img = resize_to_640(roi)
    cv2.imwrite(DEBUG_INPUT, inp_img)

    inp = inp_img[..., ::-1][None, ...]

    outputs = rknn.inference([inp])

    # ---------- 2️⃣ proto ----------
    proto = outputs[12][0]  # (32,160,160)

    best_score = -1
    best_coef = None

    out_i = 0
    for stride in STRIDES:
        reg = outputs[out_i][0]
        cls = outputs[out_i + 1][0, 0]
        obj = outputs[out_i + 2][0, 0]
        coef = outputs[out_i + 3][0]
        out_i += 4

        score_map = sigmoid(cls) * sigmoid(obj)
        y, x = np.unravel_index(np.argmax(score_map), score_map.shape)
        score = score_map[y, x]

        if score < OBJ_THRESH or score <= best_score:
            continue

        best_score = score
        best_coef = coef[:, y, x]

    if best_coef is None:
        return None

    # ---------- 3️⃣ proto_mask（完整） ----------
    proto_mask = sigmoid(np.tensordot(best_coef, proto, axes=1))  # (160,160)

    pm = (proto_mask - proto_mask.min()) / (proto_mask.max() - proto_mask.min() + 1e-6)
    cv2.imwrite(DEBUG_PROTO, (pm * 255).astype(np.uint8))

    # ---------- 4️⃣ 二值化 + 最大连通域（不裁！） ----------
    inst_proto = (proto_mask > MASK_THRESH).astype(np.uint8)
    inst_proto = largest_cc(inst_proto)

    cv2.imwrite(DEBUG_INST_PROTO, inst_proto * 255)

    # ---------- 5️⃣ proto → ROI ----------
    inst_roi = cv2.resize(
        inst_proto, (w0, h0), interpolation=cv2.INTER_NEAREST
    )

    return inst_roi * 255


# ======================
# 主程序
# ======================
def main():
    img = cv2.imread(IMAGE_PATH)
    overlay = img.copy()

    rknn = RKNNLite()
    rknn.load_rknn(MODEL_PATH)
    rknn.init_runtime()

    for (x, y, w, h) in ROIS:
        roi = img[y:y + h, x:x + w]
        mask = infer_single_roi(rknn, roi)
        if mask is None:
            continue

        color = np.zeros_like(roi)
        color[mask == 255] = (0, 255, 0)

        overlay[y:y + h, x:x + w] = cv2.addWeighted(
            roi, 0.7, color, 0.3, 0
        )

    rknn.release()
    cv2.imwrite(OUT_OVERLAY, overlay)
    print("✅ 完成:", OUT_OVERLAY)


if __name__ == "__main__":
    main()