Compression of YOLOX object detection network and deployment 
on FPGA

Qiang Cheng; Yong Bai; Lu Chen

doi:10.1117/12.2629145

22 April 2022 Compression of YOLOX object detection network and deployment on FPGA

Qiang Cheng, Yong Bai, Lu Chen

Proceedings Volume 12174, International Conference on Internet of Things and Machine Learning (IoTML 2021); 121740J (2022) https://doi.org/10.1117/12.2629145
Event: International Conference on Internet of Things and Machine Learning (IoTML 2021), 2021, Shanghai, China

Abstract

Object detection is an important task in computer vision. There are many practical applications using object detection based on deep learning nowadays. For deployment on FPGA with limited resource and operator support, object detection faces problems such as how to improve speed and reduce power consumption. YOLOX is a high-performance anchor-free YOLO version. To deploy YOLOX network on FPGA, we first replace the Focus layer of YOLOX, adjust the structure of the SPP layer, and change the activation function to meet the operator support constraints. Then we perform sparse training and use scaling factors of BN layer to select out the insignificant channels. The convolutional layer channels are pruned according to the degree of sparseness and pruning ratios. Finally, the network is quantified, compiled via the Vitis AI tool, and deployed on the Xilinx FPGA development board. Comparing the performance with different pruning ratios, the experiments demonstrate that the network runs significantly faster on the FPGA after pruning, and the power consumption is also reduced.

Citation Download Citation

Qiang Cheng, Yong Bai, and Lu Chen "Compression of YOLOX object detection network and deployment on FPGA", Proc. SPIE 12174, International Conference on Internet of Things and Machine Learning (IoTML 2021), 121740J (22 April 2022); https://doi.org/10.1117/12.2629145

ACCESS THE FULL ARTICLE

INSTITUTIONAL
Select your institution to access the SPIE Digital Library.

SELECT YOUR INSTITUTION

PERSONAL
Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.

PERSONAL SIGN IN

No SPIE Account? Create one

PURCHASE THIS CONTENT

SUBSCRIBE TO DIGITAL LIBRARY

50 downloads per 1-year subscription

Members: $195

Non-members: $335 ADD TO CART

25 downloads per 1 - year subscription

Members: $145

Non-members: $250 ADD TO CART

PURCHASE SINGLE ARTICLE

Includes PDF, HTML & Video, when available