目标检测图像数据增强（Data Augmentation）——对比度|加噪|随机调整颜色

数据扩增

数据扩增是扩充数据的一项重要手段，原理是基于已有的标注数据，进行相应变化，其对应的标签不作改变或者也跟着有相应的变化。

本次只对原始图片进行更改，label保持不变。即只变动原始数据，生成新的图像的label与原始图像一样，只是文件名变了

import skimage
import os
import shutil
import sys
from PIL import Image, ImageEnhance, ImageChops
import random
import numpy as np


def aj_contrast(root_path, img_name, label_path):  # 调整对比度 两种方式 gamma/log
    image = skimage.io.imread(os.path.join(root_path, img_name))
    gam = skimage.exposure.adjust_gamma(image, 0.5)
    img_save_path = "/xxxx/image/"
    label_save_path = "/xxxx/label/"
    skimage.io.imsave(os.path.join(img_save_path, img_name.split('.')[0] + '_gam.jpg'), gam)
    label_src = label_path + img_name.split('.')[0] + '.txt'
    label_dst = label_save_path + img_name.split('.')[0] + '_gam.txt'
    shutil.copy(label_src, label_dst)

    log = skimage.exposure.adjust_log(image)
    img_save_path = "/xxxx/image/"
    label_save_path = "/xxxx/label/"
    skimage.io.imsave(os.path.join(img_save_path, img_name.split('.')[0] + '_log.jpg'), log)
    label_dst = label_save_path + img_name.split('.')[0] + '_log.txt'
    shutil.copy(label_src, label_dst)
    return gam, log


def randomGaussian(root_path, img_name, label_path, mean=0, sigma=4):  # 高斯噪声
    image = Image.open(os.path.join(root_path, img_name))
    im = np.array(image)
    # 设定高斯函数的偏移
    means = 0
    # 设定高斯函数的标准差
    sigma = 25
    # r通道
    r = im[:, :, 0].flatten()
    # g通道
    g = im[:, :, 1].flatten()
    # b通道
    b = im[:, :, 2].flatten()

    # 计算新的像素值
    for i in range(im.shape[0]*im.shape[1]):

        pr = int(r[i]) + random.gauss(0, sigma)

        pg = int(g[i]) + random.gauss(0, sigma)

        pb = int(b[i]) + random.gauss(0, sigma)

        if(pr < 0):
            pr = 0
        if(pr > 255):
            pr = 255
        if(pg < 0):
            pg = 0
        if(pg > 255):
            pg = 255
        if(pb < 0):
            pb = 0
        if(pb > 255):
            pb = 255
        r[i] = pr
        g[i] = pg
        b[i] = pb
    im[:, :, 0] = r.reshape([im.shape[0], im.shape[1]])

    im[:, :, 1] = g.reshape([im.shape[0], im.shape[1]])

    im[:, :, 2] = b.reshape([im.shape[0], im.shape[1]])
    gaussian_image = Image.fromarray(np.uint8(im))
    img_save_path = "/xxxx/image/"
    label_save_path = "/xxxx/label/"
    gaussian_image.save(os.path.join(img_save_path, img_name.split('.')[0] + '_randomGaussian.jpg'))
    label_src = label_path + img_name.split('.')[0] + '.txt'
    label_dst = label_save_path + img_name.split('.')[0] + '_randomGaussian.txt'
    shutil.copy(label_src, label_dst)
    return gaussian_image


def randomColor(root_path, img_name, label_path):  # 随机颜色
    """
    对图像进行颜色抖动
    :param image: PIL的图像image
    :return: 有颜色色差的图像image
    """
    image = Image.open(os.path.join(root_path, img_name))
    random_factor = np.random.randint(0, 31) / 10.  # 随机因子
    color_image = ImageEnhance.Color(image).enhance(random_factor)  # 调整图像的饱和度
    random_factor = np.random.randint(10, 21) / 10.  # 随机因子
    brightness_image = ImageEnhance.Brightness(color_image).enhance(random_factor)  # 调整图像的亮度
    random_factor = np.random.randint(10, 21) / 10.  # 随机因子
    contrast_image = ImageEnhance.Contrast(brightness_image).enhance(random_factor)  # 调整图像对比度
    random_factor = np.random.randint(0, 31) / 10.  # 随机因子
    redu_image = ImageEnhance.Sharpness(contrast_image).enhance(random_factor)  # 调整图像锐度

    img_save_path = "/xxxx/image/"
    label_save_path = "/xxxx/label/"
    redu_image.save(os.path.join(img_save_path, img_name.split('.')[0] + '_redu.jpg'))
    label_src = label_path + img_name.split('.')[0] + '.txt'
    label_dst = label_save_path + img_name.split('.')[0] + '_redu.txt'
    shutil.copy(label_src, label_dst)
    return None


def process(root_path, img_name, label_path):
    # aj_contrast
    aj_contrast(root_path, img_name, label_path)
    # randomGaussian
    randomGaussian(root_path, img_name, label_path)
    # randomColor
    randomColor(root_path, img_name, label_path)


if __name__ == '__main__':
    image_path = "/xxxx/image/"
    label_path = "/xxxx/label/"
    imgs_file = sys.argv[1]
    with open(imgs_file) as f:
        for ln in f:
            img_name = ln.rstrip('\n')
            process(image_path, img_name, label_path)

上述代码是批量处理的，如果是单张处理，直接更改最后的几行代码即可

总结

本代码没有添加翻转等操作，这是因为该代码适用于目标检测数据集，目标检测数据集旋转时候，其对应的label也需要进行变换，旋转的数据增强见目标检测数据增强—-旋转

感谢阅读！