Multi-Grained Similarity Preserving and Updating for Unsupervised Cross-Modal Hashing

Round 1

Reviewer 1 Report

The authors propose a novel unsupervised hashing learning framework to improve cross-modal hashing performance. The methodology is suitable and the results are clearly presented. I recommend the authors to add Discussion section in which to compare their results with those of other studies. It will be good if the authors state the limitation of their study.

Minor editing of English language required

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

The article investigates the multi-grained similarity preserving and updating for unsupervised cross-modal hashing. After the introduction, the article presents a literature review. After that, the authors present the methodology. Section 4 contains the experiments: the authors tested the proposed algorithm with two benchmark datasets. The last section is the conclusion. I have the following comments:

Literature review: the literature review section presents many related articles. The detail of the literature review is adequate. In order to improve the presentation, I recommend adding a summary table, which contains, for example, the following header: reference, authors, algorithm, and problem.

Methodology: the authors present the methodology in detail. Notations and their meaning tables are also used for mathematical formulas. Figures and tables serve to make it easier to understand the methodology part. The only thing I miss from this part is the abbreviations and their meaning table.

Experiments: The authors tested the proposed algorithm on benchmark data. In addition, it was compared with nine other methods, based on which the proposed algorithm is sufficiently compared and its effectiveness is proven. Are these nine other methods the implementation of the authors? There is also no information on the running times of the algorithms. It is a special positive that the authors also performed hyperparameterization. I also recommend a more detailed presentation of the software.

Conclusion: this part is very short. The achieved results should be summarized in more detail. This section could be extended by summarizing the experiments section. The future work is also missing.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

The paper presents a method called MGSPU that addresses challenges in unsupervised cross-modal hashing. Cross-modal hashing is a technique used to represent data from different sources (like images and texts) in a common space for efficient retrieval.

The challenges addressed by MGSPU are:

1. Accuracy Issue: Existing methods struggle to achieve accurate cross-modal similarity learning. MGSPU proposes a new method to update a similarity matrix, which helps improve accuracy by removing noise in the original cross-modal features.

2. Preserving Similarity Structure: The current methods often neglect global and local similarity structures. MGSPU introduces a novel method for preserving multi-grained similarity, aiming to enhance the learning of cross-modal hash codes by maintaining consistency in similarity structures.

The paper claims to comprehensively verify the proposed method through experiments on two widely used datasets, comparing its performance against nine state-of-the-art competitors. According to the results, MGSPU demonstrates superior performance in cross-modal hashing.

The manuscript is well-organised, and the authors' contributions are well-specified in the beginning. The methodology is presented with a good and extensive description and the mathematical proof seems sound.

I am a little bit concerned regarding the too-technical language that the authors use, especially in the abstract. It could be difficult for readers to clearly understand the content of the paper, so I'd suggest simplifying a little the text.

Also, the conclusion section is quite scarce. Authors should provide a better insight on their findings, limitations, and potential future works.

I find the article quite difficult to read, both because of the very high level of technical language used by authors, and some weird phrasing. There is no major problem with the overall quality, it is just very complex sometimes

Author Response

Please see the attachment.

Author Response File: Author Response.pdf