数据增强

数据增强目的是人工扩展样本，提高模型鲁棒性，降低过拟合风险，其方法有：random erasing、 cutout、 hide-and-seek、grid mask、Adversarial Erasing、mixup、cutmix、mosaic、Stylized-ImageNet、label smooth和dropblock。

1、random erasing

 

"""
re=w/h
s=w*h=re*h*h=w*w/r
h=sqrt(s/r)
w=sqrt(s*r)
"""

def random_erase(img,p=0.8,sl=0.02,sh=0.5,re=0.3):
    if random.uniform(0,1)>p:
        return img

    img_area=img.shape[0]*img.shape[1]
    while True:
        erase_area=random.uniform(sl,sh)*img_area
        we,he=int(math.sqrt(erase_area*re)),int(math.sqrt(erase_area/re))
        xe=random.randint(0,img.shape[1])
        ye=random.randint(0,img.shape[0])

        if we+xe<=img.shape[1] and ye+he<=img.shape[0]:
            img[ye:ye+he,xe:xe+we,:]=np.random.randint(0,255,size=(he,we,img.shape[2]))
            return img

img=cv2.imread("D:/data/testpic/1.png")
print(img.shape[0],img.shape[1])
img1=random_erase(img.copy())
cv2.imshow('test',img1)
cv2.waitKey()

2、CutMix

https://github.com/clovaai/CutMix-PyTorch

3、mosaic数据增强

目的将4张图片合成一张图片并整合标签

首先创建一个空的画布，在某一范围内随机选择一个点作为mosaic图片的中心点（xc，yc），以该中心点可将画布分为4块

其中各块的坐标表示如下：

考虑到4块区域与4张图宽高不匹配，可能需要截取图中一部分并合成。

合成后的标签框相对位置改变。通过计算宽高（pw，ph）在合成后图片的标签框相对位置。

import os
import cv2
import numpy as np
import random
def load_img_label(trainpath):
    global img4, y1a, y2a, x1a, x2a, y1b, y2b, x2b, x1b
    with open(trainpath,'r') as f:
        imgpath=[line.strip() for line in f.readlines()]

    labelpath=[x.replace('images','labels').replace('jpg','txt').replace('png','txt') for x in imgpath]
    filename=[os.path.splitext(os.path.split(fn)[1])[0] for fn in imgpath]

    print(imgpath)
    print(labelpath)
    print(filename)

    for i in range(0,len(filename),4):
        file=filename[i:i+4]
        imgp=[]
        labels=[]
        for j in file:
            imgp.append([p for p in imgpath if os.path.splitext(os.path.split(p)[1])[0]==j][0])
            lb=[p for p in labelpath if os.path.splitext(os.path.split(p)[1])[0]==j][0]
            with open(lb,'r') as f:
                label=[line.strip().split(' ') for line in f.readlines()]
                # label=

            labels.append(label)
        #
        s = 416
        xc,yc=[int(random.uniform(s*0.5,s*1.5)) for _ in range(2)]

        for i in range(4):
            img=cv2.imread(imgp[i])
            H,W=img.shape[:2]

            if i==0:
                img4 = np.full((s * 2, s * 2, img.shape[2]), 114, dtype=np.uint8)

                x1a,y1a,x2a,y2a=max(xc-W,0),max(yc-H,0),xc,yc
                #原图截取区域
                x1b, y1b, x2b, y2b = W-(x2a-x1a), H-(y2a-y1a),W,H
            elif i == 1:
                x1a, y1a, x2a, y2a = xc, max(yc - H, 0), min(xc+W,s*2), yc
                x1b, y1b, x2b, y2b = 0, H - (y2a - y1a), min(W, x2a - x1a), H
            elif i == 2:
                x1a, y1a, x2a, y2a = max(xc - W, 0), yc, xc, min(yc+H,s*2)
                x1b, y1b, x2b, y2b = W - (x2a - x1a), 0, max(xc, W), min(y2a - y1a, H)
            elif i == 3:
                x1a, y1a, x2a, y2a =  xc, yc,min(xc+W,s*2), min(yc+H,s*2)
                x1b, y1b, x2b, y2b = 0, 0, min(W, x2a - x1a), min(y2a - y1a, H)

            img4[y1a:y2a,x1a:x2a]=img[y1b:y2b,x1b:x2b]
            #相对于原图位置是增加还是减少
            padw=x1a-x1b
            padh=y1a-y1b
            for lb in labels[i]:
                # for box in lb:
                x,y,w,h=float(lb[1]),float(lb[2]),float(lb[3]),float(lb[4])
                x1,y1=int((x-w/2)*W)+padw,int((y-h/2)*H)+padh
                x2, y2 = int((x + w / 2) * W)+padw, int((y + h / 2) * H)+padh
                cv2.rectangle(img4, (x1, y1), (x2, y2), (255, 0, 255), 1)


        cv2.imshow('test',img4)
        cv2.imwrite('./res.jpg',img4)
        cv2.waitKey()

        # print(labels)


# imgpath='D:/data/Mydata/images'
trainpath='D:/data/Mydata/train.txt'
load_img_label(trainpath)

以上存在一些缺点，有些合成后的标签框只显示部分，应该尽量将这部分剔除。合成后的图片还需要进行翻转、缩放、色域变换等操作。

未完待续。。。

参考：

https://zhuanlan.zhihu.com/p/41679153

https://www.cnblogs.com/monologuesmw/p/12932407.html

https://blog.csdn.net/qq_39523365/article/details/112525776?utm_medium=distribute.pc_relevant.none-task-blog-baidujs_title-2&spm=1001.2101.3001.4242