Compressing YOLO network by compressive sensing

Wu, Yirui; Meng, Zhouyu; Palaiahnakote, Shivakumara; Lu, Tong

Abstract

Object detection is one of the fundamental challenges in pattern recognition community. Recently, convolutional neural networks (CNN) are increasingly exploited in object detection, showing their promising potentials of generatively discovering patterns from quantity of labeled images. Among CNN-based systems, we focus on one state-of-the-art architecture designed for fast object detection, named as YOLO. However, YOLO, as well as CNN-based systems are hard to deploy on embedded systems due to their computationally and storage intensive. In this paper, we propose to compress YOLO network by compressive sensing, which exploits in-herent redundancy property of parameters in layers of CNN architecture, leading to decrease the computation and storage cost. We firstly convert parameter matrix to frequency domain through discrete cosine transform (DCT). Due to the smooth property of parameters when processing images, the resulting frequency matrix are dominated by low-frequency components. Next, we prune high-frequency part to make the frequency matrix sparse. After pruning, we sample the frequency matrix with distributed random Gaussian matrix. Finally, we retrain the network to finetune the remaining parameters. We evaluate the proposed compress method on VOC 2012 dataset and show it outperforms one latest compression approach.

• Presentation Conference Type: Conference Paper (published)
• Conference Name: 2017 4th IAPR Asian Conference on Pattern Recognition (ACPR)
• Start Date: Nov 26, 2017
• End Date: Nov 29, 2017
• Online Publication Date: Dec 16, 2018
• Publication Date: Dec 16, 2018
• Deposit Date: Nov 15, 2024
• Publisher: Institute of Electrical and Electronics Engineers
• Series ISSN: 2327-0985
• Book Title: 2017 4th IAPR Asian Conference on Pattern Recognition (ACPR)
• ISBN: 9781538633557
• DOI: https://doi.org/10.1109/ACPR.2017.11