Research On Lightweight Real-Time Object Detection Algorithm Based On Neural Network

At present,most research focus on sophisticated architectures and algorithms to improve the detection accuracy.However,in the field of industrial applications,it is usually limited by computing resources and software support,complex deep learning models are not suitable or cannot be deployed in actual production environments that use mobile and edge devices.So,research on lightweight detection algorithms is of great significance for industrial applications.Therefore,this thesis analyzes the existing mainstream lightweight obeject detection algorithm,combines with advanced network structure,puts forward some innovative strategies,and builds a new object detection algorithm named Handy Det.Then carries out corresponding experiment to verify thoese previous studies.The specific work can be summarized as the following three points:1.Analyzes excellent lightweight feature extraction networks(e.g.Efficientnet-lite,Shuffle Netv2,Mobile Netv3 and Ghost Net)and summarizes the core design ideas and referential experience of them.Then,the commonly used lightweight object detection framework such as SSDLite,YOLOV4-tiny and YOLO5-s in the industrial field are analyzed.To consolidate the theoretical basis for subsequent experiments,the advantages and disadvantages of each framework are summarized.2.The structure of split inverted residuals(SIR)is designed in this thesis,it can greatly reduce parameters and floating point operations for the network.In order to compensate for the accuracy loss caused by the decrease in the number of parameters,a self-adaptive context awareness module(SACAM)is introduced to expand the receptive field of feature extraction network and improve the detection performance.Finally,the feature pyramid structure commonly used in the detection algorithm is investigated and studied,and a new bi-directional feature fusion structure named PAN-Lite is designed.In the performance experiment,SIR Unit,SACAM and PAN-Lite are used to build the detection network,and the detection head of YOLOv4-tiny is replaced by GFL.So a new object detection algorithm named Handy Det is built.The number of parameters of Handy Det is only 0.75 M,accounting for only 12.4% of YOLOv4-tiny,and the number of floating-point operations is 0.36 B,accounting for only 5.2%of YOLOv4-tiny.However,the detection accuracy of Handy Det on COCO dataset is 21.1%,only 0.6% less than that of YOLOv4-tiny.In the mobile experiment,Handy Det reaches a speed of 47.6 frames per second,far exceeds YOLOv4-tiny which only gets 30.0 frames per second.3.Investigates the label distribution strategy in the process of training neural network,and propose improvement strategies for the adaptation of Sim OTA algorithm and ultralightweight deep learning model: construct training auxiliary module AGRM to improve the output quality of training process network and the positive sample matching strategy of Sim OTA algorithm at the initial training stage was modified to screen out sufficient positive samples for the network for accelerate the network convergence.In addition,this thesis introduces the Bottleneck of Mobile Netv3 into PAN-Lite and constructs a new feature fusion structure,Mobile-PAN,which enhances the multi-scale feature fusion effect.In the performance experiment,Handy Det-Plus was constructed by adding Sim OTA+AGRM and Mobile-PAN based on Handy Det.Handy Det-Plus surpasses YOLOv4-tiny in accuracy and keeps the number of parameters under 1M.In the mobile terminal experiment,Handy Det-Plus can process 40 frames of images per second,which meets the requirements of real-time object detection.