Deep Neural Network Compression Algorithm And Its Application In Object Detection

Object detection is an important research topic in computer vision,and has been widely applied to many fields,such as intelligent transportation system,intelligent monitoring system,human-machine interaction and medical navigation surgery.In recent years,deep learning methods have shown excellent performance in many domains,especially in the field of computer vision.Currently,some object detection algorithms represented by Fast R-CNN and Faster R-CNN have initiated the craze of automatic object detection algorithm.Since most researchers focus on improving the accuracy of detection algorithms,in order to extract more complex features,the number of layers in the deep network shows an exponential growth.Nowadays,networks with hundreds of thousands of layers begin to appear.However,these detection algorithms based on deep neural networks have high requirements on the storage capacity and computation power of the hardware.Therefore,although deep neural network based object detection algorithms significant outperform traditional algorithms in accuracy,their huge storage and computational costs brings a great challenge for the deployment on resource limited platforms,such as mobile and embedded devices.Therefore,network compression,acceleration and optimization has become an urgent and important research topic both in academia and industry.Based on this requirement,a group of network compression methods appear gradually.However,most of these methods result in serious performance degradation,such as the loss of accuracy in classification tasks or the increase of miss rate in detection tasks.In this paper,we provide an overview of object detection methods and network compression methods.To explicitly address the above problems,we completed the following work in order to realize the deep neural network based object detection algorithm under resource limited platform.Firstly,we compare a variety of deep object detection methods,taking into account of the accuracy,efficiency and model size of various methods.Finally,we use the end-to-end fully convolutional neural network YOLOv2 realize pedestrian detection task and achieved state-of-the-art performance on Caltech dataset with 23.6% miss rate.Furthermore,we binarize the YOLOv2 network based on the existing network compression method BWN,thereby compressing the model by 32 times and achieving a detection rate of 66 FPS.The resulting binary network has a miss rate of 33.6%.Secondly,in view of the obvious loss of accuracy caused by the existing network compression method,we propose a novel layer-wise network binarization(LWB)method.Based on the concept of layer-wise priority,we binarize the parameters of the network in inverse order of the layer depth.Our method achieves comparable accuracy against full-precision network while resulting in 2x speed up and significant memory saving up to 32 x.It greatly suppress the loss of performance caused by network compression.Thirdly,based on our layer-wise network binarization method,we further propose a flexible partial network binarization method.We can flexibly decide whether to binarize the remaining floating layers of the network or not and explore a best trade-off between the loss of performance and the compression ratio of model.This method effectively solve the problem that the existing network compression methods can not provide the compressed network model with any accuracy.Experiments show that our algorithm can maintain the same accuracy as the full-precision network while resulting in 26 x compression ratio.In contrast,the existing network binarization algorithm BWN can only fix 32 x compression ratio,therefore,it achieves 33.6% and 18% miss rate on Caltech and INRIA dataset,respectively.However,under the same 32 x compression ratio,our proposed algorithm obtains 26.7% and 13.8% miss rate on Caltech and INRIA dataset.In conclusion,the main innovation of this paper is mainly based on the second and third work metioned above,we propose a novel and flexible network binarization method named as layer-wise binarization(LWB)algorithm.Compared with the existing methods,our algorithm can effectively suppress the loss of performance caused by network compression,and flexibily achieve the best trade-off between the loss of performance and network compression ratio.