Convolutional Neural Network Structure Design And Optimization Based On Embedded Platform

Object detection algorithm is a very active research field in the field of computer vision.Its main purpose is to identify and locate the interested objects in the scene.This detection technology has many applications in people's production and life,such as video surveillance,intelligent transportation,image retrieval,human-computer interaction and other fields.Although the current object detection algorithm based on deep learning achieves high accuracy by using large convolutional neural network,it brings huge memory occupation and computing resource consumption,which makes it difficult to deploy on edge and embedded devices with limited computing power and power consumption,and it can not realize real-time detection.In view of the above problems,this paper proposes a new algorithm based on You Only Look Once(YOLO)series of single shot target detection network design principles of lightweight target detection algorithm.This paper focuses on the research of lightweight object detection algorithm based on embedded platform.There are three main contributions.First,a lightweight convolution module NEP(No-Expansion-Projection)is designed,which uses Ghost Module replace 1x1 convolution on the basis of bottleneck structure.This module realizes information fusion and feature extraction between channels with less computational cost.The NEP module use the attention module ECA(efficient channel attention)to redistribute the weight coefficient of feature channels,which can reduce the correlation between feature channels,and improves the utilization of limited parameters of lightweight network.Second,On the basis of NEP module and the framework of YOLO series algorithm,a lightweight multi-scale object detection algorithm based on visual attention mechanism is proposed(Attentional Ghost Feature-You Only Look Once,AGF-YOLO).This algorithm solves the problem that the current large-scale deep convolutional neural network occupies a large amount of memory,requires high computing power,and can not be deployed in the edge mobile scene with limited memory.Third,based on the embedded chip Hi3516DV300 and Hisilicon intelligent vision heterogeneous acceleration platform SVP(smart vision platform),the algorithm AGF-YOLO proposed in this paper is deployed on the embedded device through the quantization and the migration of the training framework,and the real-time object detection on the embedded device is realized.Through training and testing on open data sets,the proposed lightweight object detection algorithm AGF-YOLO achieves 73.8% mAP on VOC2007 data sets,which is better than most of the current popular lightweight object detection algorithms.At the same time,the size of the parameters is only 1.54 M,and the amount of computation is only 1.6GFLOPs.When AGFYOLO is deployed on the embedded platform Hi3516DV300,the detection accuracy can reach70.52%,and the reasoning speed is 40 FPS,which meets the needs of real-time detection,and proves the effectiveness of the lightweight target detection algorithm designed in this paper.