Classification for Digital Image

This is an Matlab project for recognizing noised digital image. What we want is a programme that could automatically recognize several noised digital image. The inputs are some images like the following picture, the exception is [2, 0, 2].

中文介绍

In this project, we introduced a classification example for number images.

They are similar to Google CAPTCHAs.

There are totally 1200 images in folder 'imagedata', and ground truth in file 'labels.txt'.

The task is to make a function called my_classifier.m that decodes three digits in an image.

Environment

Matlab and Image Processing Toolbox are ready for your computer.

Run

1. New a folder named ‘labeledImage’ and run ‘DeNoise.m’
2. New a folder named ‘SplitLabeledImage’ with 3 sub-folders named ‘0’, ‘1’ and ‘2’, then run ‘Split.m’
3. Run ‘trainer.m’ to get trained network ‘net.mat’
4. Run ‘evaluate_classifier.m’

How does it work

0. An original image is

1. We firstly use threshold to convert image into binary image.
   

2. Then we use imclose to remove noises.

3. We split the denoised image into three parts. In each part, there is a single number.

4. We do the same thing for each image and save the split images into the sub-folder 0 , 1 , 2

5. We desgin a CNN like this

   layers = [
       imageInputLayer([50 35 1])
       
       convolution2dLayer(3,8,'Padding','same')
       batchNormalizationLayer
       reluLayer
       
       maxPooling2dLayer(2,'Stride',2)
       
       convolution2dLayer(3,16,'Padding','same')
       batchNormalizationLayer
       reluLayer
       
       maxPooling2dLayer(2,'Stride',2)
       
       convolution2dLayer(3,32,'Padding','same')
       batchNormalizationLayer
       reluLayer
       
       fullyConnectedLayer(3)
       softmaxLayer
       classificationLayer];

6. We write a script for test images to split them into the size that our clissifer can test. In this project, it is myclassifier.m.

7. At last, in evaluate_classifier.m, we evaluate our classifier. Note, you may see that the code in this file about the test data is our training data, but you can replace them by yourself.