On the Impact of Discrete Atomic Compression on Image Classification by Convolutional Neural Networks

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The specific comments are addressed to the authors via the attached file.

Comments for author File: Comments.pdf

Author Response

Thank you very much for taking the time and efforts to review our manuscript! Please find the detailed responses below and the corresponding corrections in the re-submitted files.

 

Comment 1: There are two figures labeled two on pages 3 and 5! The second figure should be labeled as number 3.
Response 1: Thanks a lot! We've corrected this typo (see page 6, line 209).

Comment 2: The phrase ”MobileNetV2, VGG16, VGG19, ResNet50, NASNetMobile and NASNetLarge” is repeated six times. These modes can be introduced to non-familiar readers. Then, they can be referenced to avoid the repetition.
Response 2: We appreciate for this comment! We've reduced the number of this phrase. Also, we've added a brief description of these models (see subsection 3.1, page 4).

Comment 3: The comparison of memory expenses depicted in Figure 14 should be completed. The classic mode can be added to make the full comparison.
Response 3: Thanks for this suggestion! Figure 14 has been completed, Fig. 14,c has been added (see page 15).

 

 

 

Author Response File: Author Response.doc

Reviewer 2 Report

Comments and Suggestions for Authors

The paper have studied the impact of quality loss produced by the DAC algorithm on the performance of MobileNetV2, VGG16 and VGG19, ResNet50, NASNetMobile, and NASNetLarge. The experimental results have shown that the minor negative effect on classification performance is usually produced, which means that lossy compression by the DAC algorithm preserves the principal features of an image processed.

 

1. Why did the author choose six types of CNN models? Please explain the reasons.

 

2. In section 3.3 of the experimental part, how many image samples were used when testing the classification accuracy? How were these models trained? How many samples were used for training? These details need to be explained in detail.

Author Response

Thank you very much for taking the time and efforts to review this manuscript! Please find the detailed responses below and the corresponding corrections in the re-submitted file.

 

Comment 1: Why did the author choose six types of CNN models? Please explain the reasons.
Response 1: We appreciate for this comment! We've added a brief description of these models that explains the models choice (see Subsection 3.1, page 4). The main reasons are
1) to explore models of different depth – it varies from 16 to 533;
2) to explore models with different number of parameters – it varies from 4.3M to 143.7M;
3) to analyze models that combine similarities with dissimilarities; in particular, VGG16 and VGG19 have similar architectures, but different number of layers and parameters; MobileNetV2 and NASNetMobile have low parameters count, but they are of totally different depth; MobileNetV2 and ResNet50 use residual connections, but these models have different depth and parameters count; etc.

Comment 2: In section 3.3 of the experimental part, how many image samples were used when testing the classification accuracy? How were these models trained? How many samples were used for training? These details need to be explained in detail.
Response 2: We are thankful for these comments! Test data, which are used for testing the accuracy when classifying the decompressed images, consist of 131 images of 10 classes. This is said in Subsection 3.2 (see page 5). In this research, we use neural networks with pre-trained weights. We apply TensorFlow Keras tools that provide these models with parameters trained using the ImageNet database (more than 1.2M training samples). We have presented the corresponding explanations in subsection 3.1 (see page 4).

Author Response File: Author Response.doc

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Accept

Comments on the Quality of English Language

Verify and check all English typos

Author Response

The authors have separately and carefully read the paper one more time and introduced several minor updates in the text to enhance clarity and precision. The introduced changes are marked in text.

Reviewer 2 Report

Comments and Suggestions for Authors

In the revised manuscript，the authors have addressed the concerns regarding the methodology by providing more detailed information on the experimental design  methods. The clarity of this section has been enhanced, making it easier for readers to understand and replicate the study. The authors have made substantial efforts to address the previous comments and improve the quality of their manuscript. With minor revisions to enhance clarity and precision, I believe this paper will be suitable for publication to the journal.

Author Response

The authors have separately and carefully read the paper one more time and introduced several minor updates in the text to enhance clarity and precision. The introduced changes are marked in text.