图像分割数据增强，mask重合！！！

增强主要是增加图片数量，添加噪声翻转等，mask的像素不会发生变换，但是会一起翻转，注意mask最好像素还是不要改，比如mask像素类别是0,1,2,3四个类别，就不要让他改动。本文的最后再附上检查mask像素是否超标的代码，此检查代码可以有效解决报num_classes相关的错误。

一、数据增强代码：

import imgaug.augmenters as iaa  # 导入iaa
import cv2
import glob
import os
import numpy as np

if __name__ == '__main__':
    img_dir = 'F:/CT_lung_seg_and_class/seg_data/CC-CCI/image'	# 图片文件路径
    msk_dir = 'F:/CT_lung_seg_and_class/seg_data/CC-CCI/mask'	# 标签文件路径
    #img_type = '.png'
    img_tmp_dir = 'F:/CT_lung_seg_and_class/seg_data/CC-CCI_AUG/image/'	# 输出图片文件路径
    msk_tmp_dir = 'F:/CT_lung_seg_and_class/seg_data/CC-CCI_AUG/mask/'
    img_list = os.listdir(img_dir)
    msk_list = os.listdir(msk_dir)

    for i in range(len(img_list)):
        img_name = img_list[i]
        msk_name = msk_list[i]

        img = cv2.imread(filename=img_dir + "/" + img_name)
        img = np.expand_dims(img, axis=0).astype(np.float32)
        msk = cv2.imread(filename=msk_dir + "/" + msk_name)
        msk = np.expand_dims(msk, axis=0).astype(np.int32)
        # 定义数据增强策略
        # 每次选择一个翻转方式
        seq = iaa.Sequential([
            iaa.Fliplr(0.5),    # 水平翻转
            iaa.Flipud(0.5),    # 垂直翻转
            iaa.GaussianBlur(sigma=(0, 3.0)),   # 高斯模糊
            iaa.Sharpen(alpha=(0, 0.3), lightness=(0.9, 1.1)),  # 锐化处理
            iaa.Affine(scale=(0.9, 1), translate_percent=(0, 0.1), rotate=(-40, 40), cval=0, mode='constant'),   # 仿射变换
            # iaa.CropAndPad(px=(-10, 0), percent=None, pad_mode='constant', pad_cval=0, keep_size=True), # 裁剪缩放
            # iaa.PiecewiseAffine(scale=(0, 0.05), nb_rows=4, nb_cols=4, cval=0),     # 以控制点的方式随机形变
            iaa.ContrastNormalization((0.75, 1.5), per_channel=True),  # 对比度增强，0.75-1.5随机数值为alpha，该alpha应用于每个通道
            iaa.AdditiveGaussianNoise(loc=0, scale=(0.0, 0.05 * 255), per_channel=0.5),  # 高斯噪声
            iaa.Multiply((0.8, 1.2), per_channel=0.2),  # 20%的图片像素值乘以0.8-1.2中间的数值,用以增加图片明亮度或改变颜色
        ])
        # 同时对原图和分割进行数据增强
        for j in range(8):
            img_aug, msk_aug = seq(images=img, segmentation_maps=msk)
            img_out = img_tmp_dir + img_name.split(".")[0] + "_" + str(j) + '.jpg'
            msk_out = msk_tmp_dir + msk_name.split(".")[0] + "_" + str(j) + '.png'
            cv2.imwrite(img_out, img_aug[0])
            cv2.imwrite(msk_out, msk_aug[0,:,:,0])
        print("正在进行数据增强{}".format(i))

二、将mask和图像进行加颜色重合代码：

import os
from PIL import Image

root_path_background = "F:/CT_lung_seg_and_class/seg_data/CC-CCI_clean/val/images/"
root_path_paste = "F:/CT_lung_seg_and_class/seg_data/CC-CCI_clean/val/mask/"
output_path = "F:/CT_lung_seg_and_class/seg_data/CC-CCI_clean/val/manual/"
img_list = os.listdir(root_path_background)
label_list = os.listdir(root_path_paste)
img_list = sorted(img_list)
label_list = sorted(label_list)
for num, img_label in enumerate(zip(img_list, label_list)):
    img = Image.open(os.path.join(root_path_background, img_label[0]))
    label = Image.open(os.path.join(root_path_paste, img_label[1]))
    label = label.convert("RGB")
    fin = Image.blend(img, label, 0.3)
    # fin.show()
    fin.save(os.path.join(output_path, img_label[0]))

将mask和原图经过图像重合后图片效果如下：

三、检查mask是否在限定类别以内，比如如果mask只包含0,1两个像素值，该代码可以找出超出像素1的图像。代码如下：

import os
import os.path as osp
from tqdm import tqdm
import cv2
import numpy as np


num_classes = 4
mask_dir = "F:/CT_lung_seg_and_class/seg_data/CC-CCI_clean/train/mask1/"
mask_names = os.listdir(mask_dir)

for mask_name in tqdm(mask_names):
    mask_path = osp.join(mask_dir, mask_name)
    mask = cv2.imread(mask_path, 0)
    h, w = mask.shape[:2]
    pix = []
    for i in range(0, num_classes):
        pix.append(len(np.where(mask==i)[0]))
    if sum(pix) != h*w:
        print("error: " + mask_name)