深度学习图像数据增强

---

 在做深度学习时，往往没有足够的数据去做训练，图像增强是一种较好的方法；

运用了库Albumentations

安装方法：

1. 直接pip安装；

pip --default-timeout=100 install albumentations --user

 2. 也可以直接下载压缩文件，解压到自己电脑的库环境文件夹下；

---

 label标签可以根据所需调整，本文直接用的原始的，增强可以调整参数配置；

import cv2
import random
import numpy as np
import os
import time
import albumentations as A

# 定义读取box函数 txt标签文件
def box_from_txt(txt, img_h, img_w):  
    boxes = []
    with open(txt, 'r') as f:
        for line in f.readlines():

            # print(line)
            yolo_datas = line.strip().split(' ')
            label = int(float(yolo_datas[0].strip()))
            # print(label)

    
            center_x = float(str(yolo_datas[1]))
            center_y = float(str(yolo_datas[2]))
            center_width = float(str(yolo_datas[3]))
            center_height = float(str(yolo_datas[4]))

            loc = [label,center_x,center_y,center_width,center_height]
            boxes.append(loc)
    return boxes


# 定义图像处理函数
transform = A.Compose([
    #A.RandomCrop(width=256, height=256),
    #A.HorizontalFlip(p=0.5),
    A.RandomBrightnessContrast(brightness_limit=(0.1,0.5), contrast_limit=(0.1,0.5),p=0.2),
    #
    #A.JpegCompression(quality_lower=19, quality_upper=20, p=1),
    A.Blur(blur_limit=(15, 15), p=1),
    #
    A.RGBShift(r_shift_limit=(10,50), g_shift_limit=(10,50), b_shift_limit=(10,50), p=0.3),
    
    A.MultiplicativeNoise(multiplier=(0.3, 1.1), p=0.5),
   
    A.ColorJitter(brightness=(0.3,0.7),contrast=(0.3,0.7),saturation=(0.3,0.7),hue=(-0.5,0.5),p=0.3),
    
    A.Posterize(num_bits=2, p=1)
    
    A.MotionBlur(blur_limit=(5,25), p=0.1),
    
    A.RandomGamma(p=1),
    
    A.Downscale(scale_min=0.25, scale_max=0.75, p=1),
    
    A.RandomShadow(shadow_roi=(0, 0, 1, 1), p=0.1),
])



# 文件路径
imgfile = r'./new_img/'   
txtfile = r'./new_txt/'   
savepath = r'./out/'
img_list = os.listdir(imgfile)
txt_list = os.listdir(txtfile)


for i in range(0,10000):
    index_img = np.random.randint(0, len(img_list))
    imgpath = imgfile + img_list[index_img]
    #print(imgpath)
    txtpath = txtfile + txt_list[index_img]
    name_img = imgpath.split('/')[-1].split('.')[0]
    name_txt = txtpath.split('/')[-1].split('.')[0]

    img = cv2.imdecode(np.fromfile(imgpath, dtype=np.uint8), -1)
    img_h, img_w = img.shape[0], img.shape[1]
    boxes = box_from_txt(txtpath, img_h, img_w)
    out_file = open(savepath + 'out_%04d.txt' % (i), 'w', encoding='UTF-8')

    #time.sleep(2)
    #print(boxes)

    [ out_file.write(" ".join(  [str(j) for j in ii]   ) + '\n') for ii in boxes ]


    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)


    transformed = transform(image=img)
    transformed_image = transformed["image"]
    transformed_image = cv2.cvtColor(transformed_image, cv2.COLOR_RGB2BGR)

    cv2.imwrite(savepath + 'out_%04d.jpg' % (i), transformed_image)


print("save successful!")