现在有好多做图像增强的，直接拿来用就好了，这个是自己写的记录一下

import numpy as np
import cv2

# 剪切一张图片
def crop_2D(inputImg, label, crop_size=[256, 256]):
    index_arr = np.where(label == 1)
    # if len(index_arr[0]) == 0:
    #     return inputImg, label
    choice_index = np.random.randint(len(index_arr[0]))
    index_x = index_arr[0][choice_index]
    index_y = index_arr[1][choice_index]

    start_x = index_x - crop_size[0] // 2
    start_y = index_y - crop_size[1] // 2
    end_x = index_x + crop_size[0] // 2
    end_y = index_y + crop_size[1] // 2

    # 上下方向越界
    if start_x < 0:
        end_x = end_x - start_x
        start_x = 0
    elif end_x > crop_size[0] * 2:
        start_x = start_x - (end_x - crop_size[0] * 2)
        end_x = crop_size[0] * 2

    # 左右方向越界
    if start_y < 0:
        end_y = end_y - start_y
        start_y = 0
    elif end_y > crop_size[1] * 2:
        start_y = start_y - (end_y - crop_size[1] * 2)
        end_y = crop_size[1] * 2

    inputImg1 = inputImg[start_x: end_x, start_y: end_y]
    label1 = label[start_x: end_x, start_y: end_y]
    return inputImg1, label1

# 尺度放大
def zoom_2D(inputImg, label):
    src_w, src_h = inputImg.shape
    src_w_half = src_w // 2
    src_h_half = src_h // 2

    zoom = np.random.uniform(0.825, 1.125)
    if zoom > 1:
        interpolation = cv2.INTER_CUBIC
    elif zoom < 1:
        interpolation = cv2.INTER_AREA
    inputImg = cv2.resize(src=inputImg, dsize=(0, 0), fx=zoom, fy=zoom, interpolation=interpolation)
    label = cv2.resize(src=label, dsize=(0, 0), fx=zoom, fy=zoom, interpolation=interpolation)
    if zoom > 1:
        w, h = inputImg.shape
        inputImg1 = inputImg[w // 2 - src_w_half: w // 2 + src_w_half, h // 2 - src_h_half: h // 2 + src_h_half]
        label1 = label[w // 2 - src_w_half: w // 2 + src_w_half, h // 2 - src_h_half: h // 2 + src_h_half]
    elif zoom < 1:
        inputImg1 = np.zeros([src_w, src_h])
        label1 = np.zeros([src_w, src_h])
        w, h = inputImg.shape

        if w % 2 != 0:
            inputImg1[src_w_half - w // 2: src_w_half + 1 + w // 2, src_h_half - h // 2: src_h_half + 1 + h // 2] = inputImg[:, :]
            label1[src_w_half - w // 2: src_w_half + 1 + w // 2, src_h_half - h // 2: src_h_half + 1 + h // 2] = label[:, :]
        else:
            label1[src_w_half - w // 2: src_w_half + w // 2, src_h_half - h // 2: src_h_half + h // 2] = label[:, :]
            inputImg1[src_w_half - w // 2: src_w_half + w // 2, src_h_half - h // 2: src_h_half + h // 2] = inputImg[:, :]

    return inputImg1, label1


# 角度
def rot_2D(inputImg, label, rng=np.random.RandomState(1)):
    angle = np.random.randint(360)
    rows, cols = inputImg.shape
    M = cv2.getRotationMatrix2D(((cols - 1) / 2.0, (rows - 1) / 2.0), angle, 1)
    # perform the actual rotation and return the image
    return cv2.warpAffine(inputImg, M, (cols, rows)), cv2.warpAffine(label, M, (cols, rows))

# 翻转
def flip_2D(inputImg, label):
    if np.random.randint(2):
        # 水平翻转
        inputImg = np.fliplr(inputImg)
        label = np.fliplr(label)
    else:
        # 垂直翻转
        inputImg = np.flipud(inputImg)
        label = np.flipud(label)
    return inputImg, label