【图像脏污检测】

---

前言

图像脏污检测，可利用亮度差和卷积神经网络等检测方法。

---

一、Blemish check

将图像分块检测每块与周围的亮度差.
图像均值、方差、标准差 :
均值(mean) = (a0, a1…an-1) / n
方差= ((a0 - mean)² + (a1 - mean)² +…+ (an-1 - mean)²) / n
标准差 = 方差开方
标准差反映了图像像素值与均值的离散程度。

二、检测步骤

1.引入库

from PIL import Image, ImageStat
import cv2
import numpy as np
import matplotlib.pyplot as plt

2.检测图像

代码如下：

class blemish(object):
        '''
        This module is for blemish check:
                Global blemish check
                Regional blemish check            
        '''
        def __init__(self):
             print('\n******blemish check******')
        def Regional_blemish(self):
            '''
                Regional blemish check:
                1,divide image into M*N blocks;
                2,get average luminance and standard deviation of each block. If Std>0.5*Average, mark the block as "NG".
                3,check each block. Skip the "NG" block. Get the max delta between a block with its adjacent 8 blocks. 
                4,Get the max delta.        
            '''
            #img = Image.open('blemish.png')
            width, height = img.size
            # block size
            block_size_w = int(width/16)#64
            block_size_h = int(height/12)#48
            list_mean=[]# get luminance of each block
            list_std=[]

            for x in range(0, width, block_size_w):
                for y in range(0, height, block_size_h):
                    x1 = x
                    y1 = y
                    x2 = x + block_size_w
                    y2 = y + block_size_h
                    # get block
                    block = img.crop((x1, y1, x2, y2))
                    # save block
                    block.save('{}_{}_{}.jpg'.format(x1, y1, block_size_w))
                    # get the mean and std
                    blocks=block.convert('L')# 
                    blocks=blocks.getdata()
                    blocks=np.array(blocks) 
                    print('Block average luminance and standard deviation: %.2f ,%.2f'%(blocks.mean(0),blocks.std(0)))
                    list_mean.append(blocks.mean(0))
                    list_std.append(blocks.std(0))
                    if blocks.std(0)>0.5 * blocks.mean(0):
                        print('\tThe %d_%d block is NG,Skip.\n'%(x1,y1))
                    else:
                        #Get the delta
                        delta=[]
                        # 4 Neighborhood or 8 neighborhood
                        coordinate=[(x1-block_size_w,y1,x2-block_size_w, y2),(x1,y1-block_size_h,x2, y2-block_size_h),(x1+block_size_w,y1,x2+block_size_w, y2),(x1,y1+block_size_h,x2, y2+block_size_h)]
                        for i in coordinate:
                            block_1= img.crop(i)
                            block_1=block_1.convert('L').getdata()
                            block_1=np.array(block_1) 
                            delta_3=blocks.mean(0)-block_1.mean(0)
                            delta.append(delta_3)
                        print('The delta of the 4 neighborhood:%s'%delta)
                        deltal_2=max(delta)# get the max value among them
                        print('The max delta: %.2f'%deltal_2)
                        threshold_2=10
                        if deltal_2>threshold_2:
                            print('\nWarning "Blemish Detected!\n"')
                        else:
                            print('\nGoto OTP write\n')
                        pass

原始图片：

有脏污的块：
Block average luminance and standard deviation: 2.89 ,26.98
The 80_0 block is NG,Skip.
Block average luminance and standard deviation: 31.75 ,84.20
The 80_53 block is NG,Skip.
Block average luminance and standard deviation: 0.84 ,14.63
The 120_0 block is NG,Skip.
Block average luminance and standard deviation: 9.26 ,47.71
The 120_53 block is NG,Skip.

---

总结

本文利用图像均值和标准差检测白色脏污，作简单展示，完整的脏污检测算法有待进一步训练。