Adaptive Neural-Network-Based Lossless Image Coder with Preprocessed Input Data
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThis paper proposes a lossless image encoding method based on Artificial Neural Networks (ANN), with a particular focus on enhancing the performance of the ANN model through preprocessing the input data. By preprocessing the input images, the paper demonstrates how to improve data compression performance while reducing computational complexity. However, this paper lacks innovation and primarily suffers from the following issues:
1. In Section 6, "Preprocessing Data for AdNN," the description of the proposed concepts and methods is overly concise. A complete processing procedure, underlying principles, and relevant background information should be supplemented.
2. The comparative methods are outdated and should be supplemented with two recent approaches from recent years.
3. The experimental results are limited and should be supplemented with visual experiments, otherwise it is insufficient to fully validate the conclusions of this paper.
4.The method proposed in this paper is a data preprocessing approach. Here, only bpp (bits per pixel) is provided, but other indicators such as PSNR (Peak Signal-to-Noise Ratio) and MS-SSIM (Multi-Scale Structural Similarity Index) are not given, which does not effectively demonstrate the superiority of the method. Other specific evaluation metrics should be supplemented.
Author Response
- In Section 6, "Preprocessing Data for AdNN," the description of the proposed concepts and methods is overly concise. A complete processing procedure, underlying principles, and relevant background information should be supplemented.
- Hopefully the extended text of section 5 clarifies the proposed method. Pixel estimator is provided at the output of the rightmost neuron in Figure 2, its inputs come from the hidden layer to the left, equations from (3) to (8). In Concept 1 input nodes are fed with values defined in Table 1, input to Concept 2 are predictions provided by ALCM, CoBALP, RLS, and LA-OLS algorithms. Descriptions of these methods appeared in other papers, hence, testing software may treat their repetition as auto-plagiarism.
- The comparative methods are outdated and should be supplemented with two recent approaches from recent years.
- Table 3 contains comparison of our methods to the newest deep learning ones: LCIC, L3C, CWPLIC, and LCIC duplex. On the other hand, Table 2 shows progress that was achieved with respect to older AdNN codecs, and two classic ones: JPEG-LS, and CALIC. The achievements of deep learning techniques are obtained not without a cost. The topic is covered in section 2, deeper analysis can be found in our previous paper:
- Ulacha, M. Łazoryszczak, Lossless Image Compression Using Context-Dependent Linear Prediction Based on Mean Absolute Error Minimization, Journal: Entropy (2024), ISSN 1099-4300, pp. 1?23, vol. 26, no. 12, paper no. 1115. https://doi.org/10.3390/e26121115
- The experimental results are limited and should be supplemented with visual experiments, otherwise it is insufficient to fully validate the conclusions of this paper.
- The decoded images are identical to the originals, these are lossless image coding techniques. Figure 5 shows to what extent the coded images genres varied.
4.The method proposed in this paper is a data preprocessing approach. Here, only bpp (bits per pixel) is provided, but other indicators such as PSNR (Peak Signal-to-Noise Ratio) and MS-SSIM (Multi-Scale Structural Similarity Index) are not given, which does not effectively demonstrate the superiority of the method. Other specific evaluation metrics should be supplemented.
- The decoded images are identical to the originals, these are lossless image coding techniques. For example, the PSNR value is infinity.
Author Response File: Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThe authors proposed a lossless image compression technique based on Adaptive NN. Compared to classical approaches such JPEG-LS, the pixel prediction is done using ANN. Two versions of the technique are presented, where the input to the ANN are different for two cases. Result shows better compression rate using their proposed approach. However, the improvement is marginal and there is huge cost of delay as encoding and decoding is very slow.
Below are some comments
1. The title should be “Adaptive Neural Network Based Lossless Image Coder with Preprocessed Input Data”, instead of “Artificial Neural Network …”
2. In line 76-77, authors need to clarify how W = 3 corresponds to Fig. 3. Clarify what is meant by “W pixels up and to the left from P(0) plus W-pixel extension above to the right”
3. In the result section, include the encoding and decoding time for each test image, and compare with classical method such as JPEG-LS.
4. In line 6, do you mean “overfitting” when mentioned “overtraining”?
Author Response
1.The title should be “Adaptive Neural Network Based Lossless Image Coder with Preprocessed Input Data”, instead of “Artificial Neural Network …”
- Yes, indeed, the title is corrected.
- In line 76-77, authors need to clarify how W = 3 corresponds to Fig. 3. Clarify what is meant by “W pixels up and to the left from P(0) plus W-pixel extension above to the right”
- Hopefully the new explanation is more clear.
- In the result section, include the encoding and decoding time for each test image, and compare with classical method such as JPEG-LS.
- Coding time depends solely on the number of pixels, its independent on image contents. This means that coding times for e.g. Lennagray and Peppers are the same. In conclusion a note is added: “Coding time for JPEG-LS is three orders of magnitude shorter than for methods described in the paper”.
- In line 6, do you mean “overfitting” when mentioned “overtraining”?
- Corrected
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for Authors- The fairness of the comparison with competing approaches is unclear; additional details about the parameters used for the competing methods should be provided.
- The choice of the structure of the ANN could include the architecture selection methods for multilayer feedforward networks using statistical sensitivity analysis.
- More information is needed in the paper about the step size parameter of the NLMS algorithms and its influence on the overall performances.
- There are inconsistencies in the reference section that have to be addressed, such as variations in how the authors' names are written (e.g. Eirikur Agustsson, M.T.R.T.L.V.G., Ulacha, G).
- A comparison of the considered approaches in terms of memory requirements could be beneficial for the paper.
- To promote reproducible research and have a beneficial impact on the academic community, I recommend that the authors share the associated source code on platforms like GitHub, or a similar repository.
Author Response
- The fairness of the comparison with competing approaches is unclear; additional details about the parameters used for the competing methods should be provided.
- Table 3 contains comparison of our methods to the newest deep learning ones: LCIC, L3C, CWPLIC, and LCIC duplex. On the other hand, Table 2 shows progress that was achieved with respect to older AdNN codecs, and two classic ones: JPEG-LS, and CALIC. The achievements of deep learning techniques are obtained not without a cost. The topic is covered in section 2, deeper analysis can be found in our previous paper:
- Ulacha, M. Łazoryszczak, Lossless Image Compression Using Context-Dependent Linear Prediction Based on Mean Absolute Error Minimization, Journal: Entropy (2024), ISSN 1099-4300, pp. 1?23, vol. 26, no. 12, paper no. 1115. https://doi.org/10.3390/e26121115
- One of goals of the paper is to show that initial preprocessing of data may improve performance and time complexity of ANN based lossless codecs, possibly deep learning ones, too. Up to now nobody tried to do that.
- The choice of the structure of the ANN could include the architecture selection methods for multilayer feedforward networks using statistical sensitivity analysis.
- Definitely yes. The aim of the paper is to show that initial data preprocessing improves performance of even the simplest network architecture. Moreover, AdNNs are more universal than deep learning methods, as they need not any initial training, they learn using only data from the coded pixels surroundings.
- More information is needed in the paper about the step size parameter of the NLMS algorithms and its influence on the overall performances.
- Detailed description of the implemented NLMS are provided in the paper [12], where formula (29) shows its step size, link https://doi.org/10.3390/e22090919. The NLMS stage is used in our methods to minimize residual dependencies between distant data samples, not fully removed by the main algorithms, hence very high rank of NLMS: r=106. The approach works.
4) - There are inconsistencies in the reference section that have to be addressed, such as variations in how the authors' names are written (e.g. Eirikur Agustsson, M.T.R.T.L.V.G., Ulacha, G).
- We try our best, but there are problems with compatibility of MDPI template and BibTex. The problems are always resolved by MDPI editors, who are perfectionists in this domain.
- A comparison of the considered approaches in terms of memory requirements could be beneficial for the paper.
- An additional paragraph is added at the end of section 2: “AdNNs have much smaller memory complexity than deep learning algorithms. They contain only one hidden layer of size 12, Concept 1, or 7, Concept 2, the coder needs only minimal "cache" to remember data for them, see section 6. The deep learning networks are much larger, and may need memory for large sets of parameters.
- To promote reproducible research and have a beneficial impact on the academic community, I recommend that the authors share the associated source code on platforms like GitHub, or a similar repository.
- The current implementation of the codec has a form of typical experimental software: unclear structure, unfortunately no comments. It is far from being comprehensive for an outsider. The clear version of the software is being written by a Ph.D. student, it will be generally available.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsThe authors have answered all my concerns. I have no other questions.
Reviewer 3 Report
Comments and Suggestions for AuthorsThe authors have sufficiently addressed my comments.