语义分割的labe与分类问题不同，语义分割的label是一个mask，所以训练图像在做增广的时候，mask也要做相应的变换。
如：原来的图片和mask如下：
做旋转后：
做翻转后：
所以对于语义分割而言，其数据增广是连同mask一起做的。
---------------------------------------------------------------------------------------------------
以下提供一份代码，集成了 torchvision.transforms.functional 里的比较常用的数据增广方法(包括旋转，上下翻转，左右翻转，裁剪，调整对比度，调整饱和度，调整亮度，中心裁剪等)：
此代码用于在训练前通过对原来的数据做多种转换生产更多的训练数据。
此代码完全可以直接调用，用一个.py文件存放即可。
代码：
import random
import os
import numpy as np
from PIL import Image
from torchvision import transforms
import torchvision.transforms.functional as tf
class Augmentation:
def __init__(self):
pass
def rotate(self,image,mask,angle=None):
if angle == None:
angle = transforms.RandomRotation.get_params([-180, 180]) # -180~180随机选一个角度旋转
if isinstance(angle,list):
angle = random.choice(angle)
image = image.rotate(angle)
mask = mask.rotate(angle)
image = tf.to_tensor(image)
mask = tf.to_tensor(mask)
return image, mask
def flip(self,image,mask): #水平翻转和垂直翻转
if random.random()>0.5:
image = tf.hflip(image)
mask = tf.hflip(mask)
if random.random()<0.5:
image = tf.vflip(image)
mask = tf.vflip(mask)
image = tf.to_tensor(image)
mask = tf.to_tensor(mask)
return image, mask
def randomResizeCrop(self,image,mask,scale=(0.3,1.0),ratio=(1,1)):#scale表示随机crop出来的图片会在的0.3倍至1倍之间，ratio表示长宽比
img = np.array(image)
h_image, w_image = img.shape
resize_size = h_image
i, j, h, w = transforms.RandomResizedCrop.get_params(image, scale=scale, ratio=ratio)
image = tf.resized_crop(image, i, j, h, w, resize_size)
mask = tf.resized_crop(mask, i, j, h, w, resize_size)
image = tf.to_tensor(image)
mask = tf.to_tensor(mask)
return image, mask
def adjustContrast(self,image,mask):
factor = transforms.RandomRotation.get_params([0,10]) #这里调增广后的数据的对比度
image = tf.adjust_contrast(image,factor)
#mask = tf.adjust_contrast(mask,factor)
image = tf.to_tensor(image)
mask = tf.to_tensor(mask)
return image,mask
def adjustBrightness(self,image,mask):
factor = transforms.RandomRotation.get_params([1, 2]) #这里调增广后的数据亮度
image = tf.adjust_brightness(image, factor)
#mask = tf.adjust_contrast(mask, factor)
image = tf.to_tensor(image)
mask = tf.to_tensor(mask)
return image, mask
def centerCrop(self,image,mask,size=None): #中心裁剪
if size == None:size = image.size #若不设定size，则是原图。
image = tf.center_crop(image,size)
mask = tf.center_crop(mask,size)
image = tf.to_tensor(image)
mask = tf.to_tensor(mask)
return image,mask
def adjustSaturation(self,image,mask): #调整饱和度
factor = transforms.RandomRotation.get_params([1, 2]) # 这里调增广后的数据亮度
image = tf.adjust_saturation(image, factor)
#mask = tf.adjust_saturation(mask, factor)
image = tf.to_tensor(image)
mask = tf.to_tensor(mask)
return image, mask
def augmentationData(image_path,mask_path,option=[1,2,3,4,5,6,7],save_dir=None):
'''
:param image_path: 图片的路径
:param mask_path: mask的路径
:param option: 需要哪种增广方式：1为旋转，2为翻转，3为随机裁剪并恢复原本大小，4为调整对比度，5为中心裁剪(不恢复原本大小)，6为调整亮度,7为饱和度
:param save_dir: 增广后的数据存放的路径
'''
aug_image_savedDir = os.path.join(save_dir,'img')
aug_mask_savedDir = os.path.join(save_dir, 'mask')
if not os.path.exists(aug_image_savedDir):
os.makedirs(aug_image_savedDir)
print('create aug image dir.....')
if not os.path.exists(aug_mask_savedDir):
os.makedirs(aug_mask_savedDir)
print('create aug mask dir.....')
aug = Augmentation()
res= os.walk(image_path)
images = []
masks = []
for root,dirs,files in res:
for f in files:
images.append(os.path.join(root,f))
res = os.walk(mask_path)
for root,dirs,files in res:
for f in files:
masks.append(os.path.join(root,f))
datas = list(zip(images,masks))
num = len(datas)
for (image_path,mask_path) in datas:
image = Image.open(image_path)
mask = Image.open(mask_path)
if 1 in option:
num+=1
image_tensor, mask_tensor = aug.rotate(image, mask)
image_rotate = transforms.ToPILImage()(image_tensor).save(os.path.join(save_dir, 'img', str(num) + '__rotate.jpg'))
mask_rotate = transforms.ToPILImage()(mask_tensor).save(os.path.join(save_dir, 'mask', str(num) + '_rotate_mask.jpg'))
if 2 in option:
num+=1
image_tensor, mask_tensor = aug.flip(image, mask)
image_filp = transforms.ToPILImage()(image_tensor).save(os.path.join(save_dir,'img',str(num)+'_filp.jpg'))
mask_filp = transforms.ToPILImage()(mask_tensor).save(os.path.join(save_dir,'mask',str(num)+'_filp_mask.jpg'))
if 3 in option:
num+=1
image_tensor, mask_tensor = aug.randomResizeCrop(image, mask)
image_ResizeCrop = transforms.ToPILImage()(image_tensor).save(os.path.join(save_dir, 'img', str(num) + '_ResizeCrop.jpg'))
mask_ResizeCrop = transforms.ToPILImage()(mask_tensor).save(os.path.join(save_dir, 'mask', str(num) + '_ResizeCrop_mask.jpg'))
if 4 in option:
num+=1
image_tensor, mask_tensor = aug.adjustContrast(image, mask)
image_Contrast = transforms.ToPILImage()(image_tensor).save(os.path.join(save_dir, 'img', str(num) + '_Contrast.jpg'))
mask_Contrast = transforms.ToPILImage()(mask_tensor).save(os.path.join(save_dir, 'mask', str(num) + '_Contrast_mask.jpg'))
if 5 in option:
num+=1
image_tensor, mask_tensor = aug.centerCrop(image, mask)
image_centerCrop = transforms.ToPILImage()(image_tensor).save(os.path.join(save_dir, 'img', str(num) + '_centerCrop.jpg'))
mask_centerCrop = transforms.ToPILImage()(mask_tensor).save(os.path.join(save_dir, 'mask', str(num) + '_centerCrop_mask.jpg'))
if 6 in option:
num+=1
image_tensor, mask_tensor = aug.adjustBrightness(image, mask)
image_Brightness = transforms.ToPILImage()(image_tensor).save(os.path.join(save_dir, 'img', str(num) + '_Brightness.jpg'))
mask_Brightness = transforms.ToPILImage()(mask_tensor).save(os.path.join(save_dir, 'mask', str(num) + '_Brightness.jpg'))
if 7 in option:
num+=1
image_tensor, mask_tensor = aug.adjustSaturation(image, mask)
image_Saturation = transforms.ToPILImage()(image_tensor).save(os.path.join(save_dir, 'img', str(num) + '_Saturation.jpg'))
mask_Saturation = transforms.ToPILImage()(mask_tensor).save(os.path.join(save_dir, 'mask', str(num) + '_Saturation.jpg'))
augmentationData(r'E:\datasets\isbi\train\images',r'E:\datasets\isbi\train\label',save_dir=r'E:\代码\mytext\suanfa\aug')
运行效果：
