Project/Area Number |
23K16861
|
Research Category |
Grant-in-Aid for Early-Career Scientists
|
Allocation Type | Multi-year Fund |
Review Section |
Basic Section 60040:Computer system-related
|
Research Institution | Yokohama National University |
Principal Investigator |
孫 鶴鳴 横浜国立大学, 大学院工学研究院, 准教授 (90835886)
|
Project Period (FY) |
2023-04-01 – 2026-03-31
|
Project Status |
Granted (Fiscal Year 2023)
|
Budget Amount *help |
¥4,680,000 (Direct Cost: ¥3,600,000、Indirect Cost: ¥1,080,000)
Fiscal Year 2025: ¥650,000 (Direct Cost: ¥500,000、Indirect Cost: ¥150,000)
Fiscal Year 2024: ¥1,820,000 (Direct Cost: ¥1,400,000、Indirect Cost: ¥420,000)
Fiscal Year 2023: ¥2,210,000 (Direct Cost: ¥1,700,000、Indirect Cost: ¥510,000)
|
Keywords | Image Compression / 画像圧縮 / 深層学習 |
Outline of Research at the Start |
Learned image compression (LIC) has shown a superior coding ability than the traditional image compression standards. However, current LIC is still not practical due to several problems. This research aims at a practical LIC system by the co-development of both algorithm and architecture.
|
Outline of Annual Research Achievements |
This year, I mainly worked on two topics. First is the algorithm-architecture co-optimization of learned image compression (LIC) on FPGA. Based on a pipelined architecture, the input and output channel parallelism is restricted to ease the routing phase, so that more DSP can be used. After that, the neural network channel is searched to improve the DSP efficiency. With high DSP utilization and efficiency, compared with the recent work, the throughput can be improved by at most 1.5x. Besides, the compression efficiency will not be affected after the neural network search. Second is the privacy of LIC for machine. The overall framework includes the client side which captures image and the cloud side which conducts the machine vision. To avoid the privacy leakage in the cloud side, feature of captured image is generated at the client side, coded by an LIC autoencoder and sent to the cloud side. The cloud side then decodes the feature and performs the machine vision based on the decoded feature. By optimizing the layer number of feature, we can not only reach a good trade-off between rate for the feature transmission and accuracy for the machine vision, but also avoid leaking the privacy information in the cloud.
|
Current Status of Research Progress |
Current Status of Research Progress
2: Research has progressed on the whole more than it was originally planned.
Reason
This year, I have presented one domestic conference at IPSJ. In addition, one paper at IEEE international conference ISCAS and one paper at IEEE Journal on Emerging and Selected Topics in Circuits and Systems have been accepted. The domestic conference and journal paper are about the algorithm-architecture co-optimization of LIC. ISCAS paper is about the privacy of LIC on the usage of machine vision. Therefore, the research is undergoing well as scheduled.
|
Strategy for Future Research Activity |
There are two plans. First is to realize the variable bitrate (VBR). Currently, multiple LIC models are required to support VBR, which is very memory consuming. To solve this problem, the most efficient way is to add a gain unit which is learnable through the training. However, adding the gain unit cannot realize an effective result on the high bitrate. To solve this problem, we plan to add the gain unit not only in the bottleneck layer, but also in the intermediate layers. Second is to improve the robustness of LIC network. Since the attacker can use FGSM to generate the adversarial samples. The bit consumption will become larger or the reconstruction quality will become worse. To solve this problem, we plan to use adversarial training to include the adversarial samples in the training.
|