A Novel FPGA-Based Architecture for Fast Automatic Target Detection in Hyperspectral Images

Round 1

Reviewer 1 Report

In this manuscript, authors propose a novel FPGA based fast implementation of the Automatic Target Generation Process based on Orthogonal Subspace Projector (ATGP-OSP), denoted as Fast-ATGP. This fast implementation is expected to allow on-board real-time operations in hyperspectral remote sensing, without drop in the accuracy. Two major optimizations are included, briefly summarized as: (i) replacing complex matrix inversion procedures by a fixed operation scale and (ii) replacing the huge matrix multiplication during projection by proper vectorization.

The proposal is implemented using High Level Synthesis (HLS) with an efficient architecture on a Xilinx XC7VX690T. Two datasets are used in the experiments: Cuprite from AVIRIS and Urban from HYDICE. The detection accuracy is evaluated in terms of Spectral Angle Mapper (SAM). Comparing ATGP-OSP with the Fast-ATGP proposed implementation, similar SAM values are obtained for both datasets, and for the Cuprite data, a speed up of 34x is achieved, where the proportion of resources is much reduced for BRAMs, DSPs, and LUTs (although increased for Registers). Other results include comparing the processing time of the proposed Fast-ATGP for both C++ and FPGA, and also a small analysis on the trend of speedup and proportion of resources when varying the number of Processing Elements (PE).

I find the paper well written and well structured, the proposal seems reasonable and worth of publication. I just have some issues to be addressed:

1) Section 4.1 seems a little bit incomplete. Part of section 5.1 should be included here for clarity.

2) Are the numbers of Figure 9 based on any dataset?

3) Minor typos, such as: (i) PE acronym should be included in the abbreviations section, (ii) in line 395 ‘Table7’ should be ‘Table 7’, (iii) ‘spectral angel mapper’ should be ‘spectral angle mapper’ across the whole document.

Author Response

Dear Editor and Reviewers,

Thank you for your thoughtful review of our manuscript entitled “A Novel FPGA-based Architecture for Fast Automatic Target Detection in Hyperspectral Images” (ID: 406251).  We take your concerns seriously and have addressed them point by point. Detailed responses and remarks are provided in the additional PDF file.

Best regards,

Jie Lei

Author Response File: Author Response.pdf

Reviewer 2 Report

This paper presents a new method for HSI detection. This method is an amelioration of ATGP-OSP one by integrating the parallel process. The proposed method is implemented on FPGA which support the real time systems.

The paper is well written. However, some modification/clarification should be added :

1) For algorithm 1, how to know the number of targets to detect ? It is given by the ground-truth in the databases ? If yes, this clarification should be added

2) In line 2 of algorithm 1, I think you must change x0 by m0

3) For the comparison, the proposed method is compared with one method. How about other method in the literature ?

4) A visual comparison of the detected targets can be a fruitful information for the readers.

5) How about the GPU ? It is already tested ? This issue can be considered as a futur work of this contribution

Author Response

Dear Editor and Reviewers,

Thank you for your thoughtful review of our manuscript entitled “A Novel FPGA-based Architecture for Fast Automatic Target Detection in Hyperspectral Images” (ID: 406251).  We take your concerns seriously and have addressed them point by point. Detailed responses and remarks are provided in the additional PDF file.

Best regards,

Jie Lei

Author Response File: Author Response.pdf

Reviewer 3 Report

This paper presents a FPGA framework for hyper spectral imaging. The paper  provides a hardware method for target detection in hyperspectal images. Some comments are below

 1. There are some parts of the paper which is not explained clearly. For example, Algorithm 1 on page 4 describes the main steps of an algorithm. However, some variables are not clearly presented. For example, which does variable m_0 stand for? Please provide explanations. In this context, the respective text indicates (line 120 on page 3 ); it finds a target m_1.... linearly spanned by m_0!!!!! 

2. The same problems exist in the whole paper. For example, algorithm 2 table on page 7 is very difficult to follow. Please provide reference to Equation and simplify the writing style. 

3. I will recommend to provide an visual example of a hyper spectral image indicating which is the targets, (is that spatial regions in a hyper spectral image of particular spectral signatures) and in this section to clarify the symbols. Again I recommend to indulge a nomenclature before introduction. 

4. Recently there are some papers proposing deep learning for hyper spectral analysis and categorisation. How these methods are related with the one proposed in this paper. I mean the proposed hardware implementation can exploit deep learning implementation scenarios. Can this method apply to deep learning architectures?

Refs

The three most cited papers in deep learning and hyperspectral (from scopus) 

Chen, Y., Lin, Z., Zhao, X., Wang, G., Gu, Y.

Deep learning-based classification of hyperspectral data(2014) IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 7 (6), art. no. 6844831, pp. 2094-2107. Cited 423 times.

Zhao, W., Du, S. 

Spectral-Spatial Feature Extraction for Hyperspectral Image Classification: A Dimension Reduction and Deep Learning Approach (2016) IEEE Transactions on Geoscience and Remote Sensing, 54 (8), art. no. 7450160, pp. 4544-4554. Cited 133 times.

Makantasis, K., Karantzalos, K., Doulamis, A., Doulamis, N. 

Deep supervised learning for hyperspectral data classification through convolutional neural networks (2015) International Geoscience and Remote Sensing Symposium (IGARSS), 2015-November, art. no. 7326945, pp. 4959-4962. Cited 124 times.

Author Response

Dear Editor and Reviewers,

Thank you for your thoughtful review of our manuscript entitled “A Novel FPGA-based Architecture for Fast Automatic Target Detection in Hyperspectral Images” (ID: 406251).  We take your concerns seriously and have addressed them point by point. Detailed responses and remarks are provided in the additional PDF file.

Best regards,

Jie Lei

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

I’m happy with the updates in the manuscript. Hence, I think the paper can be accepted. Just one small thing:

1) Please make sure the explanation about Fig. 10 (it applies to all datasets) is included in the final manuscript.

Author Response

Dear Editor and Reviewers,

Thank you for your thoughtful review of our manuscript entitled “A Novel FPGA-based Architecture for Fast Automatic Target Detection in Hyperspectral Images” (ID: 406251). We appreciate for your warm work earnestly. Detailed response is provided in the additional PDF file.

Best regards,

Jie Lei

Author Response File: Author Response.pdf

Reviewer 2 Report

This paper can be published in this revised form

Author Response

Dear Editor and Reviewers,

Thank you for your thoughtful review of our manuscript entitled “A Novel FPGA-based Architecture for Fast Automatic Target Detection in Hyperspectral Images” (ID: 406251). Thank you very much for your approval of our work, which is an encouragement and a motivation for us. We appreciate for your warm work earnestly, and thank you a lot.

Best regards,

Jie Lei

Reviewer 3 Report

The Authors have addressed all my previous comments. 

Author Response

Dear Editor and Reviewers,

Thank you for your thoughtful review of our manuscript entitled “A Novel FPGA-based Architecture for Fast Automatic Target Detection in Hyperspectral Images” (ID: 406251). Thank you very much for your approval of our work, which is an encouragement and a motivation for us. We appreciate for your warm work earnestly, and thank you a lot.

Best regards,

Jie Lei