Neural Network System for Recognizing Images Affected by Random-Valued Impulse Noise

Round 1

Reviewer 1 Report

Recognition of images is a hot topic and still can't be treated as well solved task (despite huge progress made in the last decade). The authors are paying main attention to the problem of lowering the effect of noise (in particular impulse noise) on the accuracy of image recognition. This is a significant problem and advances in this area are very useful. So the paper is important and may be interesting for a relatively big audience. 

Together it is necessary to provide some remarks, comments and observations. The Chapter is called Materials and Methods. The content of the chapter is related to the most popular image recognition and noise processing methods. Remains unclear what the authors have in mind using the term Materials. 

Chapter 3.1 introduces the proposed method. It is highly desirable to provide a deeper explanation of how and why the authors proposed the current method. What could be the reason why the previous studies not used noise detection?

It is necessary to provide more details on how the authors added the noise to the images. In particular, what does the 1,10,25,50 percent mean? Is this number of pixels that were corrupted by noise or the strength of the impulse itself? Even impulse noise typically corrupts some part of the image. How the size of the area affected by impulse was evaluated in the study? How the strength of the impulse has been simulated?

Some English editing is necessary too because some sentences aren't very clear. E.g. in line authors are stating that "CNN are the most optimal technology for pattern recognition". It is unclear what the most optimal means. As well as for different tasks different methods are better suited. Maybe the authors had in mind image recognition tasks only? In several places (e.g. in line 94) Cyrillic symbols were left. More clearly needs to be presented the start of Chapter 3.1. In the current form, the reader gets the impression that the authors are proposing a system of neural networks. The essential proposal is to implement a noise identifier and apply denoising if the noise has been detected. My suggestion is to state this very clearly.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

This paper proposed a workflow to detect and clean noisy images to improve NN classification accuracy. In general, the method proposed is technically sound and well-presented. However, I hope the authors could explain some confusion in this paper.

1. In lines 264 to 267, the authors wrote “During the experiment, it was found that for noise with intensity n=0.1 the best result is at threshold T=22, for n =0.25 the best result is at threshold T=19, for n=0.5 the best result is at threshold T=100. Thus, at high noise intensity, the need to increase the threshold value is determined ……”. Though the authors draw a conclusion that a higher threshold value is needed for higher noise, it is not supported by the numbers reported. N=0.25 is higher than n=0.1 but T=19 is smaller than T=22. Could you explain?

2. Figure 5-8 reported the results from N=0.01 to N=0.5, could the authors also report the T for N=0.01? In reality, the noise level may not be known, and multiple noise levels may exist. Could the authors explain how to handle known and multiple noise level cases?

3. In lines 243 to 247, the authors presented the formula for determining the coefficients. However, the meanings of some parameters are not explicitly defined (e.g. k, l, a, b, etc.). Could the authors define them explicitly in the paragraphs followed?

 

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

1.      The paper introduced a three-stage image cleaning and recognition system, which includes a developed detector of pulsed noisy pixels, a filter for cleaning found noisy pixels based on an adaptive median, and a neural network program for recognizing cleaned images. Neither the motivations nor the contributions are well described. Many researchers have worked on image denoising using various filters and on image recognition using CNN model. The question is what is your novelty in this research?

2.      You put everything in the introduction section. You have to separate the introduction (problem definition, problem statement, and motivations) from related works, which include a related sequence of research, and determine the field gap among researchers. Consequently, you proposed a model or technique to solve that problem. Honestly, all those points are missing.

3.      You used CIFAR10 dataset. Is it balanced? Did you try to make balance between classes? Try to augment your data to reduce also the over-fitting issue.  

4.      Figure 1 does not reflect the architectural steps of CNN. It is only presents the process diagram of the work. It is quite interesting to investigate the hyper-parameters of CNN model such as (Learning rate, batch size, epoch no., and convolution filter size) in order to show their effects on the recognition rates.    

5.      Load noisy images from dataset and apply median filter to remove or reduce the impact of Random-Valued Impulse Noise is a simple process while changing the noise rate. Explain how did you select the noise percentage? And how affect the threshold values?

6.      I did not understand what it means Table 1? Is it comparison with your proposed model?

7.      Figure 3 is not required.

8.       In Figure 4, a comparison of CIFAR10 image recognition results cleaned from 25% random valued impulse noise is presented. However, you should present the optimal noise ratio, which is 10% in your work. Another point is that the confusion matrix must have the same number of image in all cases. But in figure 4, each CM has different image numbers.

9.      How did you split your data for the training and testing? Did CM present the trained or tested data?

10.  It would be better to draw the accuracy and loss curves versus epoch no. for train and test validation. 

11.  Did you consider the accuracy comparison of your proposed model and others related works?

Comments for author File: Comments.docx

Author Response

Please see the attachment

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

The authors took into the account majority of the remarks

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

1. Add subsection 1.2 Research Motivations starting from line to the end of section 1.

2. I did not satisfy with the response of question 4. Please reconsider it.

Author Response

Please see the attachment.

Author Response File: Author Response.docx