语义分割实现地表建筑物识别2 数据扩增部分

语义分割实现地表建筑物识别2 数据扩增部分

主要内容

• 学习常用数据扩增方法，并用opencv和albumentations库完成具体操作
• 学习pytorch读取图像

数据扩增的意义：增加训练样本量，也可以有效增加样本的语义空间。是一种有效的正则化方法，可以防止模型过拟合。
数据扩增方法分类：

• 标签不变的数据扩增方法：数据变换后图像类别不变
• 标签变化的数据扩增方法：数据变换后图像类别变化。
  对于语义分割，常规的数据扩增方法都是标签变化的扩增方法。
  这次学习的数据扩增方法主要是垂直/水平翻转、裁剪、旋转等。

课后作业

1. 使用OpenCV完成图像加噪数据扩增

import cv2
import numpy as np 
import pandas as pd
import matplotlib.pyplot as plt 
import albumentations as A

def rle_decode(mask_rle,shape=(512,512)):
    '''将rle格式进行解码为图片'''
    if mask_rle is np.nan:
        return None
    s = mask_rle.split()
    starts,lengths = [np.asarray(x,dtype=int) for x in (s[0:][::2],s[1:][::2])]
    starts -= 1
    ends = starts + lengths
    img = np.zeros(shape[0]*shape[1],dtype=np.uint8)
    for lo,hi in zip(starts,ends):
        img[lo:hi] = 1
    return img.reshape(shape,order='F')

def fig_plot(img1,img2):
    plt.figure(figsize=(16,8))
    plt.subplot(1,2,1)
    plt.imshow(img1)

    plt.subplot(1,2,2)
    plt.imshow(img2)

原始图片

train_mask = pd.read_csv('./train_mask.csv',sep='\t',names=['name','mask'])

i = 23557
img = cv2.imread('./train/'+train_mask['name'].iloc[i])
mask = rle_decode(train_mask['mask'].iloc[i])
fig_plot(img,mask)

翻转

# 垂直翻转
fig_plot(cv2.flip(img,0),cv2.flip(mask,0))
# 水平翻转
fig_plot(cv2.flip(img,0),cv2.flip(mask,0))

裁剪

# 随机裁剪
x,y = np.random.randint(0,256),np.random.randint(0,256)

fig_plot(img[x:x+256,y:y+256],mask[x:x+256,y:y+256])

• https://stackoverflow.com/questions/14435632/impulse-gaussian-and-salt-and-pepper-noise-with-opencv#https://stackoverflow.com/questions/14435632/impulse-gaussian-and-salt-and-pepper-noise-with-opencv#
• https://www.cnblogs.com/wmy-ncut/p/9995687.html
• https://www.cnblogs.com/mlan/p/8179028.html

添加噪声

'''
Parameters
----------
image : ndarray
    Input image data. Will be converted to float.
mode : str
    One of the following strings, selecting the type of noise to add:

    'gauss'     Gaussian-distributed additive noise.
    'poisson'   Poisson-distributed noise generated from the data.
    's&p'       Replaces random pixels with 0 or 1.
    'speckle'   Multiplicative noise using out = image + n*image,where
                n,is uniform noise with specified mean & variance.
'''

import numpy as np
import os
import cv2

def noisy(noise_typ,image):
    if noise_typ == "gauss":
        row,col,ch= image.shape
        mean = 0
        std = 20 #直接影响噪声质量，数值越大噪声比例越高
        gauss = np.random.normal(mean,std,(row,col,ch))  
        gauss = gauss.reshape(row,col,ch)
        noisy_ = image + gauss
        return noisy_.astype('uint8')
        # noisy = image.copy()
        # gauss = np.zeros(image.shape, np.uint8) # do not use original image it overwrites the image
        # mean = 0
        # sigma = 100 #100的时候偏红是什么情况
        # gauss = cv2.randn(gauss,mean,sigma) # create the random distribution
        # noisy = cv2.add(noisy, gauss) # add the noise to the original image
        # return noisy
    elif noise_typ == "s&p":
        s_vs_p = 0.5
        amount = 0.004
        out = image.copy()

        # Salt mode
        num_salt = np.ceil(amount * image.size * s_vs_p)
        coords = [np.random.random_integers(0, i - 1, int(num_salt))
                for i in [image.shape[0],image.shape[1]]]  #注意第三个channel要全部置位，否则会变成红绿色的椒盐 啊哈哈哈
        out[coords] = 255

        # Pepper mode
        num_pepper = np.ceil(amount* image.size * (1. - s_vs_p))
        coords = [np.random.random_integers(0, i - 1, int(num_pepper))
                for i in [image.shape[0],image.shape[1]]]  #注意第三个channel要全部置位，否则会变成红绿色的椒盐 啊哈哈哈
        out[coords] = 0
        return out
    elif noise_typ == "poisson":
        vals = len(np.unique(image))
        vals = 2 ** np.ceil(np.log2(vals))
        noisy = np.random.poisson(image * vals) / float(vals)
        return noisy
    elif noise_typ =="speckle":
        row,col,ch = image.shape
        gauss = np.random.randn(row,col,ch)
        gauss = gauss.reshape(row,col,ch)        
        noisy = image + image * gauss
        return noisy

1. 添加椒盐噪声

img_aug,mask_aug = augments['image'],augments['mask']

noise_gauss = noisy("s&p", img_aug) 
fig_plot(img_aug,noise_gauss)

2. 添加高斯噪声

img_aug,mask_aug = augments['image'],augments['mask']

noise_gauss = noisy("gauss", img_aug) 
fig_plot(img_aug,noise_gauss)

2. 使用OpenCV完成图像旋转数据扩增

def rotate_bound(image,angle):
    #获取图像的尺寸
    #旋转中心
    (h,w) = image.shape[:2]
    (cx,cy) = (w/2,h/2)
    
    #设置旋转矩阵
    M = cv2.getRotationMatrix2D((cx,cy),-angle,1.0)
    cos = np.abs(M[0,0])
    sin = np.abs(M[0,1])
    
    # 计算图像旋转后的新边界
    nW = int((h*sin)+(w*cos))
    nH = int((h*cos)+(w*sin))
    
    # 调整旋转矩阵的移动距离（t_{x}, t_{y}）
    M[0,2] += (nW/2) - cx
    M[1,2] += (nH/2) - cy
    
    return cv2.warpAffine(image,M,(nW,nH))
i = 23557
img = cv2.imread('./train/'+train_mask['name'].iloc[i])
img2 = rotate_bound(img,20)
fig_plot(img,img2)

3. 使用albumentations其他的操作完成扩增操作

# 水平翻转
augments = A.HorizontalFlip(p=1)(image=img,mask=mask) 
#augments = A.randomCrop(p=1,height=256,width=256)(image=img,mask=mask)
#augments = A.ShiftScaleRotate(image=img,mask=mask)
img_aug,mask_aug = augments['image'],augments['mask']
fig_plot(img_aug,mask_aug)

trfm = A.Compose([
    A.Resize(256,256),
    A.HorizontalFlip(p=0.5),
    A.VerticalFlip(p=0.5),
    A.RandomRotate90(),])
augments = trfm(image=img,mask=mask)
img_aug,mask_aug = augments['image'],augments['mask']
fig_plot(img_aug,mask_aug)

4. 使用pytorch完成赛题数据读取

• Dataset:数据集，对数据进行读取并进行数据扩增
• DataLoader:数据读取器，对Dataset进行封装并进行批量读取

import torch.utils.data as D

class TianChiDataset(D.Dataset):
    def __init__(self,paths,rles,transform):
        self.paths = paths
        self.rles = rles
        self.transform = transform
        self.len = len(paths)
    
    def __getitem__(self,index):
        img = cv2.imread(self.paths[index])
        mask = rle_decode(self.rles[index])
        augments = self.transform(image=img,mask=mask)
        return self.as_tensor(augments['image']),augments['mask'][None]

    def __len__(self):
        return self.len

IMAGE_SIZE = 256
trfm = A.Compose([
    A.Resize(IMAGE_SIZE,IMAGE_SIZE),
    A.HorizontalFlip(p=0.5),
    A.VerticalFlip(p=0.5),
    A.RandomRotate90(),
])

dataset = TianChiDataset(
    train_mask['name'].values,
    train_mask['mask'].fillna('').values,
    trfm
)

loader = D.DataLoader(
    dataset,batch_size=10,shuffle=True,num_workers=0
)

参考

• https://stackoverflow.com/questions/14435632/impulse-gaussian-and-salt-and-pepper-noise-with-opencv#https://stackoverflow.com/questions/14435632/impulse-gaussian-and-salt-and-pepper-noise-with-opencv#
• https://www.cnblogs.com/wmy-ncut/p/9995687.html
• https://www.cnblogs.com/mlan/p/8179028.html
• https://www.jianshu.com/p/9194f43fd68a