含有忍耐度的TP计算(Precision、Rcall、F1 score)

对于边缘分割类任务，边缘部分因为人为标记不准确等原因，在评估模型时，需要对这些区域设置一定忍耐度以保证模型评估更完整。这里提供代码参考。（允许两个像素差距）

import cv2
import numpy as np
import time

gt = cv2.imread("/home/harvey/Datasets/DeepCrack/OurResult/11125-1_gt.png", cv2.IMREAD_UNCHANGED)
p_msk = cv2.imread("/home/harvey/Datasets/DeepCrack/OurResult/11125-1_pred.png", cv2.IMREAD_UNCHANGED)
thres = 10
kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (5, 5))


def rolling_cal(ground, pred_msk, thres=10):
    H, W = ground.shape
    total = H * W
    filter_size = 5
    msk_t = (pred_msk > thres).astype(np.int)
    ground = (ground > 0).astype(np.int)
    H, W = ground.shape
    gt = np.pad(ground, (2, 2), 'constant', constant_values=(0, 0))
    tp, fp, fn = 0.0, 0.0, 0.0
    for r in range(0, H):
        for c in range(0, W):
            # 池化大小的输入区域
            gti = gt[r:r + filter_size, c:c + filter_size]
            cur_pexl = msk_t[r, c]
            gt_patch = (gti * kernel).sum()
            if cur_pexl != 0 and gt_patch != 0:
                tp += 1
            elif cur_pexl != 0 and gti[2][2] == 0:
                fp += 1
            elif cur_pexl == 0 and gti[2][2] != 0:
                fn += 1
    return tp, fp, fn, total

start = time.time()
tp, fp, fn, total = rolling_cal(gt, p_msk, 10)
print("time:", time.time() - start)
print(tp, fp, fn, total)
recall = tp / (tp + fn)
precision = tp / (tp + fp)
f1 = 2*tp / (2*tp + fp + fn)
print(precision, recall, f1)