数据增强 imgaug 框架使用
Posted clemente
tags:
篇首语:本文由小常识网(cha138.com)小编为大家整理,主要介绍了数据增强 imgaug 框架使用相关的知识,希望对你有一定的参考价值。
官方文档
https://imgaug.readthedocs.io/en/latest/source/overview/arithmetic.html#cutout
首先下载项目
https://github.com/aleju/imgaug
然后建立conda独立环境
conda install six numpy scipy Pillow matplotlib scikit-image imageio
conda install -c menpo opencv
代码
注意 待增强的图片需要分别放置于/abc/下单独的文件夹,以保证每个文件夹都是相关增强图片
# -*- coding: utf-8 -*- """ 图片增强 一个目录下有多个类别 """ import numpy as np import imgaug as ia import imgaug.augmenters as iaa import matplotlib.pyplot as plt import os from skimage import io import random import time #根目录,其中是不同类别的子目录,子目录中是图片 data_dir = r"/home/joe/Documents/abc/" total_count = 4 #所有子目录下文件计数 for (rootpath, dirnames, filenames) in os.walk(data_dir): for dirnametmp in dirnames: dirnametmp = data_dir+dirnametmp+"/" #每个子文件夹 for (rootpath, dirnames, filenames) in os.walk(dirnametmp): filenamescount = len(filenames) imgcounttmp = 1 #当前子目录下的文件计数 for icount in range(filenamescount): imgname = filenames.pop() #每个图片 #qqq,有imgaug的表示造的图片 file_extension = os.path.splitext(imgname) filetmp = file_extension[0] #文件名不带扩展名 extension = file_extension[1] #扩展名 if filetmp.endswith("_imgaug"): continue imgfullnametmp = dirnametmp+imgname #每个图片全路径 print( f"{imgcounttmp}/{total_count} {imgfullnametmp}") imgcounttmp += 1 total_count += 1 image = io.imread(imgfullnametmp) saveimage = [] random1 = random.uniform(-5,5) random2 = random.uniform(0,1) random3 = random.randint(1,5) augtmp = iaa.Affine(rotate=(-25+random1, 25.5+random1)) res = augtmp.augment_image(image) saveimage.append(res) augtmp = iaa.Fliplr(1.0) res = augtmp.augment_image(image) saveimage.append(res) augtmp = iaa.Crop(px=(0, random3*2), keep_size=True) res = augtmp.augment_image(image) saveimage.append(res) augtmp = iaa.Sometimes(random2,iaa.GaussianBlur(sigma=(random2, 0.5+random2))) res = augtmp.augment_image(image) saveimage.append(res) augtmp = iaa.ContrastNormalization((0.75+random2, 1.5+random2)) res = augtmp.augment_image(image) saveimage.append(res) augtmp = iaa.AdditiveGaussianNoise(loc=random3, scale=(0.01+random2, 0.08*255+random2)) res = augtmp.augment_image(image) saveimage.append(res) augtmp = iaa.Affine( scale={"x": (0.8+random2, 1.2+random2), "y": (0.8+random2, 1.2+random2)}, #缩放 translate_percent={"x": (-0.2+random2, 0.2+random2), "y": (-0.2+random2, 0.2+random2)}, #平移 rotate=(-20+random1, 60+random1), #旋转 shear=(-8+random1, 8+random1) ) res = augtmp.augment_image(image) saveimage.append(res) augtmp = iaa.GaussianBlur(sigma=random2) res = augtmp.augment_image(image) saveimage.append(res) augtmp= iaa.AverageBlur(k=random3) res = augtmp.augment_image(image) saveimage.append(res) augtmp = iaa.MedianBlur(k=random3*2+1) res = augtmp.augment_image(image) saveimage.append(res) augtmp = iaa.MotionBlur() res = augtmp.augment_image(image) saveimage.append(res) augtmp = iaa.BilateralBlur() res = augtmp.augment_image(image) saveimage.append(res) augtmp = iaa.WithColorspace(to_colorspace="HSV") res = augtmp.augment_image(image) saveimage.append(res) augtmp = iaa.AddToHueAndSaturation((-20, 20), per_channel=True) res = augtmp.augment_image(image) saveimage.append(res) augtmp = iaa.Grayscale(alpha=1.0) res = augtmp.augment_image(image) saveimage.append(res) augtmp = iaa.CLAHE(clip_limit=(1, random3*10)) res = augtmp.augment_image(image) saveimage.append(res) augtmp = iaa.GammaContrast(gamma=2*random2+1) res = augtmp.augment_image(image) saveimage.append(res) augtmp = iaa.SigmoidContrast(gain= random2+0.5) res = augtmp.augment_image(image) saveimage.append(res) augtmp = iaa.LogContrast(gain=2*random2+1) res = augtmp.augment_image(image) saveimage.append(res) augtmp = iaa.LinearContrast(alpha=random3) res = augtmp.augment_image(image) saveimage.append(res) augtmp = iaa.LinearContrast(alpha=random3*2) res = augtmp.augment_image(image) saveimage.append(res) aug = iaa.CLAHE(clip_limit=(1, random3*20)) res = augtmp.augment_image(image) saveimage.append(res) #保存图片 imgid = 0 t = time.time() tt = round(t*1000) for imgtmp in saveimage: #原文件名带有扩展名,需要拆分开 #文件名加时间戳,imgaug的表示造的图片 newname = filetmp+"_" +str(tt)+str(imgid)+"_imgaug"+extension newfullname = dirnametmp+newname #ia.imshow(imgtmp) io.imsave(newfullname,imgtmp) imgid += 1
以上是关于数据增强 imgaug 框架使用的主要内容,如果未能解决你的问题,请参考以下文章
Tensorflow 对象检测 api:如何使用 imgaug 进行增强?
目标检测数据增强:YOLO官方数据增强实现/imgaug的简单使用