BDNN: Binary convolution neural networks for fast object detection

Hanyu Peng

doi:10.1016/j.patrec.2019.03.026

BDNN: Binary convolution neural networks for fast object detection

2019-07-01

18

PRE

AI

DOI：10.1016/j.patrec.2019.03.026

Original

Original request for help

Save

Abstract

En 中文

Object detection methods based on neural networks have made considerable progress. However, methods like Faster RCNN and SSD that adopt large neural networks as the base models. It's still a challenge to deploy such large detection networks in mobile or embedded devices. In this paper, we propose a low bit-width weight optimization approach to train Binary Neural Network for object detection using binary weights in training and testing. We introduce a greedy layer-wise method to train the detection network. This method boosts the performance instead of training the entire network at the same time. Our binary detection neural network (BDNN) can reduce the computational requirements and storage with competitive performance. For example, the binary network based on Faster RCNN with VGG16 can save 95% compression. In our experiments, BDNN achieves comparable performance with mAP 63.3% and outperforms SPPNet by 4.4% on PASCAL VOC 2007. (C) 2019 Elsevier B.V. All rights reserved.

Keywords:

Deep learning

Object detection

Network compression

BDNN: Binary convolution neural networks for fast object detection

Abstract

AI Summary

Journal

Organization