Compression method based on layer-by-layer network binarization

Abstract

The invention provides a compression method based on layer-by-layer network binarization. The compression method based on layer-by-layer network binarization includes the steps: constructing a floating point type deep convolutional neural network; according to the opposite sequence of the hierarchy depth of the deep convolutional neural network, performing layer-by-layer binarization on the parameters in the network from deep to shallow until binarizing all the hierarchies in the deep convolutional neural network, and obtaining a binarized deep convolutional neural network; and performing pedestrian detection through the binarized deep convolutional neural network. Therefore, the compression method based on layer-by-layer network binarization realizes compression and acceleration of the network, and can effectively solve the problem of great precision loss caused by network quantification.

Description

technical field [0001] The invention relates to the technical field of image processing, in particular to a compression method based on layer-by-layer network binarization. Background technique [0002] Pedestrian detection aims to detect pedestrians in the image and accurately output the position and score of the candidate frame. Pedestrian detection has a wide range of applications in the field of computer vision: such as intelligent monitoring, vehicle assisted driving, intelligent robots and human behavior analysis. In recent years, with the popularity of deep learning methods, deep convolutional neural networks have become an advanced technology for solving many tasks such as pedestrian detection, pedestrian re-identification, and semantic segmentation. In order to improve the accuracy of detection, researchers generally tend to use deeper and wider neural networks. However, these convolutional neural network-based methods require a large number of floating-point oper...

AI Technical Summary

Problems solved by technology

[0004] However, due to the rough estimation method of using binary weights to approximate floating-point weights, there are estimation errors

In a deep convolutional neural network, the parameters of all layers of the network are binarized at the same time, and the estimation error is accumulated layer by layer. After passing through the multi-layer network, the error is greatly amplified, resulting in a large loss of accuracy in the detection network.

Images