[PYTHON] Data Augmentation with openCV

Introduction

How are you all doing the hot days in August?
I'm about to melt, but while listening to Kenshi Yonezu's new song </ strong> I'm coding in a room where I don't know the day and night when the air conditioner works. </ Strong>

What do you guys say summer?

• Summer festival? --Fireworks?
• shaved ice?

No, summer is the summer of *** Data Augmentation! !! !! *** *** So I will write a method of data augmentation with openCV instead of a memo.

Who is it for

--People who want to data augment a large amount of data while maintaining the directory structure! --People who get angry at someone if they don't create data quickly --People who have trouble programming! --Execute with the following command

python3 img_change.py input_directory output_directory

--Example

python3 img_change.py /data/imageset /data/imagser_change

program

img_change.py

# -*- coding: utf-8 -*-
import numpy as np

import os
import cv2
import sys
import pathlib
import glob
from scipy.signal import qspline1d, qspline1d_eval, qspline2d,cspline2d


if len(sys.argv) != 3:
    print("python3 test.py inputdir outputdir")
    sys.exit()
else:   
    input_path = sys.argv[1]
    file_list = glob.glob(input_path + "/**/**.png ", recursive=True)
    for item in file_list:
        split_name = item.split('/')
        output_name = sys.argv[2] + "/" + split_name[-3] + "/" + split_name[-2] + "/" + split_name[-1]
        output_dir = sys.argv[2] + "/" + split_name[-3] + "/" + split_name[-2]
        pathlib.Path(output_dir).mkdir(parents=True,exist_ok=True)
        img = cv2.imread(item, cv2.IMREAD_UNCHANGED)  # read image
        
        #Image processing part
        ###Increase the brightness of the image
        chg_img=img*1.2 #Brightness doubles

        
        ###Image resizing
        height = img.shape[0]
        #img.shape[0]*0.5 half the size of the original
        width = img.shape[1]
        chg_img = cv2.resize(img , (int(width), int(height)))

        ###CLAHE (redistribution so that the histogram is as evenly distributed as possible)
        clahe = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8,8))
        chg_img = clahe.apply(img)

        ###noise
        row,col,ch = img.shape
        #White
        pts_x = np.random.randint(0, col-1 , 1000) #From 0(col-1)Make a thousand random numbers up to
        pts_y = np.random.randint(0, row-1 , 1000)
        img[(pts_y,pts_x)] = (255,255,255) #y,Note that the order is x

        #black
        pts_x = np.random.randint(0, col-1 , 1000)
        pts_y = np.random.randint(0, row-1 , 1000)
        img[(pts_y,pts_x)] = (0,0,0)

        ###Distort the image
        chg_img = cv2.flip(img, 1) #Flip horizontally
        chg_img = cv2.flip(img, 0) #Invert vertically


        cv2.imwrite(output_name,chg_img )  #write image

--In the above program, the program is created with the following directory structure.

$ tree
/data
└── val
    ├── A1
    │   │  
    │   ├── 0.png
    │   ├── 1.png
    │   ├── 2.png
    │   └── ....
    ├── B1
    │   │  
    │   ├── 0.png
    │   ├── 1.png
    │   ├── 2.png
    │   └── ....    
    │            
    train
    ├── A1
    │   │  
    │   ├── 0.png
    │   ├── 1.png
    │   ├── 2.png
    │   └── ....
    ├── B1
    │   │  
    │   ├── 0.png
    │   ├── 1.png
    │   ├── 2.png
    │   └── ....    
    │              
    test
    ├── A1
    │   │  
    │   ├── 0.png
    │   ├── 1.png
    │   ├── 2.png
    │   └── ....
    ├── B1
    │   │  
    │   ├── 0.png
    │   ├── 1.png
    │   ├── 2.png
    │   └── ....    
    │            
    └── 

--Therefore, it feels like getting all .png in the directory two directories below the specified directory as a list

At the end

――There are various ways to change the image, so check it out with OpenCV image processing. --Reference site Image processing with Python (2) Data Augmentation (image padding)