[PYTHON] I made a program that solves the spot the difference in seconds

Fujitsu Systems Web Technology Advent Calendar 2019 Posted on the 16th day. The content of the article is the personal opinion, and the content of the writing is the responsibility of the author himself. The organization to which you belong does not matter.

Introduction

This time on the 16th day, I'm familiar with Saizeriya's spot the difference because I've never cleared it on my own

** A program that can take the difference between two images and display the difference in an easy-to-understand manner **

I would like to write about how to make

Rough image diagram

Implementation environment

• OS
  • MacOS Mojave10.14.5
• Language
  • python 3.7.4
• Library
  • openCV 3.4.2
  • NumPy 1.17.4
  • matplotlib 3.1.1
• Other
  • jupyter notebook 6.0.1
  • anaconda navigator

Advance preparation

(There is no guarantee, but if you can use python and various libraries, it will work normally, so please just take a look.)

• Insert a tool called anaconda navigator that can handle anaconda in an easy-to-understand manner with a gui.
• I think it's okay to install python in parallel (not necessarily in parallel).
• Use anaconda navigator to put jupyter and various libraries in it.

(Detailed installation method is'[anaconda navigator jupyter installation](https://www.google.com/search?q=anaconda+navigator+jupyter+%E3%82%A4%E3%83%B3%E3%82%B9 % E3% 83% 88% E3% 83% BC% E3% 83% AB)'I think that various things will come out.)

Source commentary

Now, here is the actual content of the source.

There are three main parts

• Create a difference mask image
• Put color on the mask part of the mask image
• Change the transparency of the mask image and combine

It feels like a mouse

import or variable setting

import cv2, matplotlib
import numpy as np
from IPython.display import Image, display_png

#Original image 1
img_01 = "sample1.png "
#Original image 2
img_02 = "sample2.png "
#Please enter the name you want to give to the output image
outImg = "output.png "

Create a difference mask image

#Read the original image 1 in color
img_src01 = cv2.imread(img_01, 1)
 
#Read original image 2 in color
img_src02 = cv2.imread(img_02, 1)
 
#Mixed normal distribution(Gaussian Mixture)It seems to be a foreground / background area division algorithm based on.
bg = cv2.bgsegm.createBackgroundSubtractorMOG()
 
#Generate mask image
mask = bg.apply(img_src01)
mask = bg.apply(img_src02)

#Output mask image
cv2.imwrite("output/machigai_mask.png ", mask)

print("Original image 1")
display_png(Image(img_01))
print("Original image 2")
display_png(Image(img_02))
print("Mask image")
display_png(Image("output/machigai_mask.png "))

The output result here is as follows.

Put color on the mask part of the mask image

mask_img = cv2.imread("output/machigai_mask.png ", 1)

#Replace the white part of the image with red
red = [240, 20, 20]
white = [255, 255, 255]
mask_img[np.where((mask_img == white).all(axis=2))] = red

# BGR to RGB
mask_img = cv2.cvtColor(mask_img, cv2.COLOR_BGR2RGB)

#Save the result.
cv2.imwrite("output/color_mask_img.png ", mask_img)

color_mask_img = cv2.imread("output/color_mask_img.png ", 1)
display_png(Image("output/color_mask_img.png "))

The output result here is as follows.

Change the transparency of the mask image and combine

If you can do this at the end, it will be completed.

# src1 =Original image 1, src2=Mask image, alpha(beta) =It seems good to adjust it at a ratio of 1 in total, gamma=It felt good when it was 0
diff_image = cv2.addWeighted(src1=img_src01,alpha=0.3,src2=color_mask_img,beta=0.7,gamma=0)
cv2.imwrite(outImg, diff_image)

display_png(Image(outImg))

Yes it's done

I tried it with an image that looks like a spot the difference

For this reason, I searched for mistakes and made it. First of all, please try it by yourself (there are 6 mistakes)

Did you find the mistake?

Next, let's take a look at the results of the fully automatic ~~ egg breaking machine ~~ solved by the spot the difference machine.

it's here

How is it? It's a great correct answer rate.

However, the detection accuracy of ** that the teacher's wig has changed to watermelon ** is not very good.

About background subtraction method

In the first place, this background subtraction method is used. This is a method for detecting moving objects on a moving image.

in short

With a fixed point camera ** Ah, people passed by ** It ’s a way to judge something like

There are two major problems when using this background subtraction method for finding mistakes.

• The size of the images to be diffed must be the same
• If the pixels are slightly out of alignment, they will all be judged differently.

So, if you try to make something that can be used for general purposes,

• Please correct a slight deviation
• Also, correct the tilt
• Correct the size
• However, recognition accuracy can be guaranteed to some extent.

I feel like I have to make such software ()

Finally

• Please correct a slight deviation
• Also, correct the tilt
• Correct the size
• However, recognition accuracy can be guaranteed to some extent.

Let's do our best to make such software next time (white eyes)

reference

-Background subtraction method -OpenCV basic knowledge -Image composition / transparency -Image fill