Group Class Residual ℓ1-Minimization on Random Projection Sparse Representation Classifier for Face Recognition
, Andriyan Bayu Suksmono 2,‡, Ian Joseph Matheus Edward 2,‡ and Koredianto Usman 3,‡
Round 1
Reviewer 1 Report
The authors have proposed two variants of the original SRC (sparse representation based classification) algorithm used for face recognition. Their main focus is to improve the computational efficiency of the original algorithm, They doen evaluations using AT&T, Extended Yale B and Georgia Tech datasets.
Overall, the paper is not is publishable state. It is full of incomplete sentences making it difficult to fathom. Authors haven’t done any effort to explain the intuition behind their ideas for these variants compared to the original algorithm nor have they reasoned about the very inconsistent result they have got. The fatal flaw of the paper is that the results don’t match the assertions and contributions the authors claim. The GC (Group class) version's accuracy is so bad compared to the original algorithm (20% vs 65%) it can’t be considered a comparable version. The other proposed version called GCR(group class residual) is slower compared to the original algorithm. So it is not computationally efficient.
Specific comments:
Line 10: expand LDA.
Sentences on lines 4-6: need to be reworded. Doesn’t make any sense.
Lines 16-18: need to be reworded.
Introduction section should have a sub-section covering different state of the art face recognition techniques.
Line 34: nonparametric learning -> provide reference for it.
Line 43-47: need to be reworded.
Create a separate section for related work.
Clearly and explicitly state the contributions of this paper and how it differs and improves from existing state-of-the-art techniques for face recognition.
Lines 112-117: can you cite empirical data or previous work about the computation efficiency issues?
Figure 3: is not referenced. You can as well remove it.
Line 164: provide references for AT&T, GT, Yale datasets.
Line 172: what do you mean by good images? Why choose only good images?
Experimental setup section is missing. No details about software/tools used for evaluation are provided.
Table 1: Provide the accuracy to at least 2 decimal digits. The whole number difference by 1 could be just rounding errors.
Table 2: provide the processing time in milli secs.
You need to compare the performance of your algorithms with current SOTA techniques on FR not just the vanilla SRC method.
Please get the paper proofread by native English speaker.
Author Response
Please see the attachment
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Dear Authors,
Excellent article on classification, with interesting results. This reviewer suggests some improvements:
- Abstract needs to modify and to be revised to be more quantitative. You can absorb readers' consideration by having some numerical results in this section.
- The discussion section in the present form is relatively weak and should be strengthened with more details and justifications.
- In conclusion section, limitations and recommendations of this research should be highlighted.
- The authors have to add the state-of-the art references in the manuscripts.
- Please leave in bold font the best results obtained in the table in both computational time and accuracy.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Reviewer 3 Report
Sorry but the current state-of-the-art in facial recognition (using convolutional networks) is much more competitive, so I fear that your article would not be of much interest to the journal's readers.
Author Response
Point 1: Sorry but the current state-of-the-art in facial recognition (using convolutional networks) is much more competitive, so I fear that your article would not be of much interest to the journal's readers.
Response 1:
Thank you for your advice on our research, perhaps we can say that the SRC technique cannot be compared with the CNN method in terms of high accuracy, but SRC will be more superior in terms of efficient computational time. Our next research paper will explore more on SRC algorithm and random projection that more efficient in computational time but still sufficiently robust on certain application, such as security of electronic transaction applications. This in general popular in term of ubiquitous computing. We will put this statement of the next development of our research on our revised paper.
Author Response File: Author Response.pdf
Reviewer 4 Report
The authors proposed a Group Class Residual ℓ1-minimization on Random Projection Sparse Representation Classifier for Face Recognition. The work is interesting and can be accepted for publication subject to minor modifications.
1. Sentence or refs like "by Wright et al." should be followed by ref. number. Check thoroughly and fix
2. Paper classification and contribution section missing in the introduction
3. Proposed algorithm needs to be elaborate. A mathematical foundation is missing.
4. Compare your results with SOTA methods. Compare with the following works
"doi: 10.1109/SPIN52536.2021.9565990"
" https://doi.org/10.1002/cpe.6157"
5. Method and result section should be more elaborative.
6. Add future works in the conclusion. What is the shortcoming of this work.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Reviewer 5 Report
Review on “Group Class Residual ℓ1-minimization on Random Projection Sparse Representation Classifier for Face Recognition”
General Description: The manuscript propose a Sparse Representation Scheme for Face Recognition. More specifically, a group-based approach is proposed. The grouping structure is applied on coefficients after solving the L1-minimization problem. The presented results fail to show the usefulness of the proposed approach. In general the manuscript need to be changed significantly. See comments below.
Comments:
1. Equation 1, with respect to the definition of matrix A in lines 88-92, is erroneous or at least very confusing. In SRC schemes, the matrix A is the dictionary matrix and not an image.
2. The regularization parameter λ is appeared for the first time in line 100. However, I can not find what this parameter regularized. It is not appear in any equation in the manuscript.
3. The use of English language is very poor. A native English speaker/writer need to check the manuscript.
4. The results are not very convincing. What cross validation technique the authors use? According to the provided results the proposed method presents slightly better accuracy in AT&T dataset, but, much worse results in the other two datasets? Why this happens? Is it due to the properties of the algorithms or due to the differences of datasets?
5. The introduction section does not provide sufficient background of the problem and many references are missing. For examples man y works in SRC using grouping structures have been proposed in the literature. The authors need to provide a comparison with them.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Round 2
Reviewer 3 Report
Your response does not change the major reason for my rejection, sorry.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Reviewer 4 Report
The authors proposed an approach using Random Projection Sparse Representation Classifier for Face Recognition. The work seems to add some contribution to the domain. I have the following observations.
1. The abstract is still not informative, discuss a bit of the result.
2. Why authors proposed a group class-based normalization even though there are various approaches can perform the task in a better way.
3. Paper classification and novelty should be added in the last paragraph in the introduction section.
4. Literature review should be a separate section. I suggest comparing the work with following works and can be referred in proposed work. "https://doi.org/10.1002/cpe.6157", "https://doi.org/10.1007/s11063-022-10818-5" "10.1109/SPIN52536.2021.9565990"
5. Rest of the method seems to be well revised and easily understandable.
6. Results need to enhance, add more sample results and elaborate the section.
7. Comparative analysis with SOTA recent method is missing,
8. very old refs need to be removed and focus on recent studies.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
