Infrared Dim and Small Target Detection Based on Background Prediction
Round 1
Reviewer 1 Report
This paper proposes a coarse-to-fine deep learning-based method is proposed to detect dim and small targets. Several concerns are summarized as follows:
1. Please further do a more comprehensive investigation on state-to-art infrared small target detection methods. To my knowledge, there are many methods based on IR image tensor also played an important role in infrared small target detection.
2. Please explain the meanings of global information, non-local correlation and local correlation in this manuscript.
3. I am confused about the two assumptions in page 2, especially the second assumption.
4. I don’t think the third contribution in this manuscript is a contribution because it is only a description of experiments result rather than the principles and innovations of your algorithm or model. Please further illustrate your contributions.
5. Please further illustrate the mechanism of Region Proposal Network (RPN) to get Ip.
6. What is the meaning of “??” in line 233?
7. In line 263, it is more proper to write “pixels” rather than “feature”. And what’s the meaning of “feature”?
8. It is not convincing to use traditional 2D-ROC curves to evaluate the performance of detection results. A 3-D ROC curve can be used to evaluate the effectiveness of a detector, target detectability, and BKG suppressibility simultaneously. Please refer to [1] and add more experimental results with 3D ROC.
9. It is not quite convincing that the superior effectiveness of your method can well estimate because there are still many clusters in the final detection maps in Figure 6.
10. Please add ablation experiments to verify each module in your algorithm.
[1] C. -I. Chang, "An Effective Evaluation Tool for Hyperspectral Target Detection: 3D Receiver Operating Characteristic Curve Analysis," in IEEE Transactions on Geoscience and Remote Sensing, vol. 59, no. 6, pp. 5131-5153, June 2021, doi: 10.1109/TGRS.2020.3021671.
A little polishing is recommended for English writing.
Author Response
Please see the attachment.
Author Response File: 
Author Response.docx
Reviewer 2 Report
This paper presents a method for detecting dim and small targets in infrared images. The proposed method utilizes a coarse-to-fine deep learning framework that combines deep learning and background prediction. The coarse detection module uses an RPN to generate masks for potential target regions. The fine detection module uses an inpainting algorithm and mask-aware dynamic filtering to refine the detection results. The inpainting algorithm predicts the background based on global semantic information, improving the accuracy of background estimation by considering mask information. Extensive experiments are conducted to demonstrate the effectiveness of the proposed architecture for detecting dim and small targets in infrared images.
I have the following comments to be addressed for the next round of reviews:
1. One of the major ideas to improve target detection is transfer learning. Consider the following works:
a. Dong, R., Xu, D., Zhao, J., Jiao, L. and An, J., 2019. Sig-NMS-based faster R-CNN combining transfer learning for small target detection in VHR optical remote sensing imagery. IEEE Transactions on Geoscience and Remote Sensing, 57(11), pp.8534-8545.
b. Wang, Z., Du, L., Mao, J., Liu, B. and Yang, D., 2018. SAR target detection based on SSD with data augmentation and transfer learning. IEEE Geoscience and Remote Sensing Letters, 16(1), pp.150-154.
c. Rostami, M., Kolouri, S., Eaton, E. and Kim, K., 2019. Deep transfer learning for few-shot SAR image classification. Remote Sensing, 11(11), p.1374.
d. Wang, P., Wang, H., Li, X., Zhang, L., Di, R. and Lv, Z., 2021. Small target detection algorithm based on transfer learning and deep separable network. Journal of Sensors, 2021, pp.1-10.
e. Xi, J., Ye, X. and Li, C., 2022. Sonar Image Target Detection Based on Style Transfer Learning and Random Shape of Noise under Zero Shot Target. Remote Sensing, 14(24), p.6260.
The above works should be discussed in the Related Work section to offer a broader perspective to the reader about existing approaches to improve target recognition.
2. In Sections 3.2 and 3.3, please add more details about the architecture of the networks, e.g., the number of layers, nodes per layer, etc.
3. I think it is a good idea to plot one of the subfigures in Figure 6 larger so more details can be clearer. I could understand Figure 6 but it took me some time to understand it.
4. The experimental results have a comparison nature which is necessary and helpful but ablative experiments are missing. I think adding 1-2 more ablative experiments can be helpful to offer a deeper insight into the algorithm. For example, the effect of each of the architectural components can be studied on the final performance.
5. Please repeat the experiments several times and report both the average performance and the standard deviation values to make the comparison statistically meaningful.
6. The GitHub link for the data is provided in the paper. Please also include the GitHub for your code so other researchers can reproduce the results.
7. Please proofread the paper well. There are ?? in the paper which shows some references in Latex are not used properly.
The English seems fine but proofreading is necessary.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Round 2
Reviewer 2 Report
My concerns have been addressed.
