A Gated Content-Oriented Residual Dense Network for Hyperspectral Image Super-Resolution
Round 1
Reviewer 1 Report
This manuscript proposes a gated content-oriented residual dense network for hyperspectral image super-resolution. However, this manuscript has the following questions and suggestions.
(1) Training details of this proposed neural network are suggested to be supplemented, including the learning rate, the epochs and some others. These parameters are important for reimplementing the experiments.
(2) According to the data in Figure 5, the parameter k is critical for the super-resolving process in the gating mechanism. If the k is constant for different scaling factors?
(3) The PSNR is calculated by the RMSE. These two metrics are supposed to exhibit the same tendency. In this way, one of them could be deleted from the tables which exhibit metric measurements.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 2 Report
Comments to the Authors
This paper proposes a hyperspectral super-resolution method based on the gated content-oriented residual dense network, in which the structure and the texture are separately processed. The research framework of this paper is generally clear, and the experimental results show that the proposed method outperform other methods. However, there are several problems need to be solved.
1. The literature review needs to be updated in a timely manner.
2. The motivation of the proposed method should be expressed clearer.
3. Please explain the role of alternating intersection, and what problems can be solved by this strategy.
4. All evaluation metrics should be explained in detail.
5. Can this gated content-oriented strategy be utilized for other mainstream networks such as diffusion models or some others?
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 3 Report
This paper proposed a gated content-oriented residual dense network (GCoRDN) for HSI super-resolution, which includes the content-oriented convolution (COC) module, the residual dense spectral attention block (RDSAB), and the gating-mechanism (GM). Experimental results and data analysis have demonstrated the effectiveness of the proposed method.
This paper is well organized, and there are some points that I am concerned as follows:
(1) In Table 1, the authors compute the average reconstruction performance only using 4 test images. Are 4 four images enough to validate the performance of the methods with different parameter configuration?
(2) The authors should make more ablation study to verify the proposed content-oriented convolution (COC) module, the residual dense spectral attention block (RDSAB) and the weight sharing attention mechanism.
(3) The authors should find some recent methods on HSI super-resolution to make a comparison.
This paper is easy to read, and there are only minor editing of English language required
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Round 2
Reviewer 3 Report
This paper has been revised and I have some minor questions.
1) In line 437 in the revised manuscript, “It can be seen that the proposed GCoRDN outperforms all the competing methods in all the five metrics” is not quite correct. On one hand, SFCSR and ERCSR perform best in some metrics. On the other hand, there are 4 metrics in Table 4. Besides, after modifying the Table 6, Table 7 and Figure 6-12, the corresponding explanation should be revised.
2) The explanation of Table 3 should be improved. The first combination corresponds to column 2. Besides, the SAM values 2.2571 and 2.2767 are better than 2.3083, please explain it.
3)Please give more details about the network structures for different variants in Table 3. For example, do the authors replace COC with another network or just remove COC?
Author Response
Please see the attachment.
Author Response File: Author Response.docx