Research On Lightweight Target Detection Method Based On Deep Learning

In the past few years,compared with traditional computer vision algorithms,deep learning has shown better results in various computer vision applications,deep learning models perform well in terms of accuracy and processing time,so based on depth Learning object detection algorithms have gradually surpassed traditional detection techniques,and with the advent of massively labeled image test datasets and breakthroughs in computing hardware,such as more powerful GPUs(graphics processing units),CPUs(central processing units),and more.With good computational procedures(libraries and frameworks),deep learning-based detection models achieve increasing accuracy on complex test datasets.However,in order to pursue high accuracy and precision,these detection models have more and more complex network structures,and the amount of parameters and calculation is too large,which is difficult to meet the use of low-power embedded devices and mobile devices,such as smart phones,video real-time monitoring devices,etc.In order to solve this problem,this paper conducts in-depth research on the network structure,receptive field enhancement,and feature fusion of the lightweight target detection model.The main research content is divided into the following two parts:(1)A detection algorithm with receptive field enhancement and convolution lightweight(RFBG-YOLO)is proposed.For the latest lightweight target detection algorithm YOLOv5 s of the YOLO series,the Bottleneck structure in the PANET feature fusion is too small,and the feature extraction is insufficient,which affects the detection accuracy.This paper proposes a multi-branch hole convolution structure RFB-Bottleneck to replace the Bottleneck structure of PANET.The optimized feature fusion method RFB-PANET can obtain a larger receptive field with fewer parameters,which is conducive to improving lightweight.The detection accuracy of the detection model.For the optimized feature fusion method RFB-PANET,although the detection accuracy has been improved,the structure has also become more complex.Therefore,Ghost Conv is introduced to replace the conventional convolution,and the increase in the computational cost of RFB-PANET is compensated by reducing the redundant calculation of the conventional convolution.The disadvantage is that the detection speed is improved and the calculation amount is reduced.(2)The RFBG-YOLO algorithm uses the multi-branch hole convolution structure RFB-Bottleneck to improve the detection accuracy,but the more complex structure affects the detection speed.A multi-bidirectional fusion lightweight target detection algorithm(Multi-TF YOLO)with shallow network and narrow channel is proposed.For the target detection network YOLOv3,the amount of parameters and computation is too large,and it is difficult to meet the needs of low-power embedded devices.In order to complete the real-time detection task with mobile devices,this paper compresses the network part of YOLOv3 to improve the inference speed of the model and greatly reduce the amount of parameters.Aiming at the problem that the deep feature map obtained by the FPN feature fusion adopted by YOLOv3 has low resolution and insufficient position information,which weakens the detection ability of large objects,this paper proposes the Multi-TF multiple bidirectional feature fusion method,which increases the automatic The bottom-up fusion path realizes that the shallow feature map utilizes the position information of the deep feature map,and this two-way feature fusion method is used as the basic feature fusion unit in the Neck part for many times,significantly improves the detection accuracy of the network,and can achieve a reasoning speed close to YOLOv5s.