随机数据增强代码

当数据集较小，或者网络过浅容易过拟合时，需要对数据集进行数据增强以丰富样本数据，让一张图片变换多种形态，增强网络的学习能力。

以下提供两种可实现代码，对数据进行批量随机变换，只需要修改输入和输出文件夹路径，即可批量生成，一张图片可扩增8倍。（默认已安装必要的包）

原始图片

原图

实现一：随机进行颜色增强，降低，修改对比度，锐度，随机水平翻转，竖直翻转，组合翻转。

import os
import random
import cv2
from PIL import Image, ImageEnhance


def operate(currentPath, filename, targetPath):
    # 读取图像
    image = Image.open(currentPath)
    image_cv = cv2.imread(currentPath)

    # 增强亮度
    enh_bri = ImageEnhance.Brightness(image)
    brightness_factor = random.uniform(0.8, 1.2)  # 随机生成0.8到1.2之间的数值
    image_brightened = enh_bri.enhance(brightness_factor)
    directory = targetPath + 'bright/'
    if not os.path.exists(directory):
        os.makedirs(directory)
    image_brightened.save(targetPath + 'bright/' + filename)  # 保存

    # 降低亮度
    enh_bri_low = ImageEnhance.Brightness(image)
    brightness_factor = random.uniform(0.5, 0.9)  # 随机生成0.5到0.9之间的数值
    image_brightened_low = enh_bri_low.enhance(brightness_factor)
    directory = targetPath + 'low_bright/'
    if not os.path.exists(directory):
        os.makedirs(directory)
    image_brightened_low.save(targetPath + 'low_bright/' + filename)

    # 改变色度
    enh_col = ImageEnhance.Color(image)
    color_factor = random.uniform(0.5, 1.5)  # 随机生成0.5到1.5之间的数值
    image_colored = enh_col.enhance(color_factor)
    directory = targetPath + 'color/'
    if not os.path.exists(directory):
        os.makedirs(directory)
    image_colored.save(targetPath + 'color/' + filename)

    # 改变对比度
    enh_con = ImageEnhance.Contrast(image)
    contrast_factor = random.uniform(0.5, 1.5)  # 随机生成0.5到1.5之间的数值
    image_contrasted = enh_con.enhance(contrast_factor)
    directory = targetPath + 'cont/'
    if not os.path.exists(directory):
        os.makedirs(directory)
    image_contrasted.save(targetPath + 'cont/' + filename)

    # 改变锐度
    enh_sha = ImageEnhance.Sharpness(image)
    sharpness_factor = random.uniform(1.5, 3.5)  # 随机生成1.5到3.5之间的数值，你可以根据需要调整这个范围
    image_sharp = enh_sha.enhance(sharpness_factor)
    directory = targetPath + 'sharp/'
    if not os.path.exists(directory):
        os.makedirs(directory)
    image_sharp.save(targetPath + 'sharp/' + filename)

    # 水平翻转
    if random.random() < 0.5:  # 以50%的概率进行水平翻转
        image_flipped_h = image.transpose(Image.FLIP_LEFT_RIGHT)
        directory = targetPath + 'flip_h/'
        if not os.path.exists(directory):
            os.makedirs(directory)
        image_flipped_h.save(targetPath + 'flip_h/' + filename)

        # 竖直翻转
    if random.random() < 0.5:  # 以50%的概率进行竖直翻转
        image_flipped_v = image.transpose(Image.FLIP_TOP_BOTTOM)
        directory = targetPath + 'flip_v/'
        if not os.path.exists(directory):
            os.makedirs(directory)
        image_flipped_v.save(targetPath + 'flip_v/' + filename)

    if random.random() < 0.5:  # 以50%的概率进行水平竖直翻转
        image_flipped_h = image.transpose(Image.FLIP_LEFT_RIGHT)
        image_flipped_v = image_flipped_h.transpose(Image.FLIP_TOP_BOTTOM)
        directory = targetPath + 'flip_hv/'
        if not os.path.exists(directory):
            os.makedirs(directory)
        image_flipped_v.save(targetPath + 'flip_hv/' + filename)

imgdir = 'your_imagepath/file/'
for parent, dirnames, filenames in os.walk(imgdir):
    for filename in filenames:
        print('filename is: ' + filename)
        # 把文件名添加到一起后输出
        imgPath = os.path.join(parent, filename)
        # 保存处理后的图片的目标文件夹
        outPath = 'save_path/file/'
        # 进行处理
        operate(imgPath, filename, outPath)

增强图片

实现二：随机水平竖直翻转，随机角度旋转，颜色变换，随机裁剪

import os
import random
import cv2
import numpy as np
from PIL import Image
from torchvision import transforms as tfs
import matplotlib.pyplot as plt

def coms(input):
    size=[200,250,300] #随机裁剪尺寸
    selected_size = random.randint(1, 3)

    im_l = tfs.Compose([
        tfs.RandomVerticalFlip(),#随机竖直翻转
        tfs.RandomHorizontalFlip(),#随机水平翻转
        #tfs.RandomRotation(45),  #随机旋转  -45 ~45
        #tfs.Resize((100, 200))          压缩
        tfs.RandomCrop((size[selected_size-1],size[selected_size-1])),
        tfs.ColorJitter(brightness=0.5, contrast=0.5, hue=0.5), #（亮度，对比度，颜色）
        #tfs.CenterCrop(300),#中心裁剪
    ])
    out = im_l(input)
    return out
if __name__ == '__main__':
    im = Image.open('your_imagepath/input.jpg')
    output_folder = 'save_path/file/'
    nrows = 3
    ncols = 3
    figsize = (8, 8)
    _, figs = plt.subplots(nrows, ncols, figsize=figsize)# 创建子图画布
    for i in range(nrows):
        for j in range(ncols):
            # 对原始图像进行变换
            transformed_image = coms(im)
            transformed_image_np = np.array(transformed_image)
            cv2.imwrite(os.path.join(output_folder, f"trans_img_{i}_{j}.png"), transformed_image_np)
            # 显示变换后的图像
            figs[i][j].imshow(transformed_image)
            figs[i][j].axes.get_xaxis().set_visible(False)
            figs[i][j].axes.get_yaxis().set_visible(False)

    plt.show()