Research And Embedded Implementation Of Lightweight Target Detection Network

Object detection is an important research direction in the field of artificial intelligence,but its detection network model is relatively complex,which requires high storage space and computing power of running devices,leading to limited application scenarios.In order to operate the target detection network on more convenient and flexible mobile edge devices and meet the requirements of practical applications,how to balance the network parameter size and hardware device restrictions to achieve fast reasoning and high-precision detection is a subject of important research value and significance.In this paper,Mobile Net V2 is used to replace the Dark Net53 feature extraction network of Yolo V3,and lightweight improvement of convolutional module is carried out on this basic network.Finally,the network is transplanted to Atlas200 developer suite to verify the effect of target detection in actual scenes.Specific work is as follows:(1)In order to effectively solve the problem of detection accuracy loss caused by reducing the number of parameters in lightweight target detection network.In this paper,a lightweight multi-target detection network MSPF-Yolo V3 is designed.By adopting multi-layer channel shuffling structure with low packet number and low memory usage,it achieves the same information exchange effect as high packet number.Moreover,the shallow features with rich location information are used as the guidance of deep features to fuse,so as to improve the detection accuracy of targets of different sizes.The m AP of the network on PASCAL VOC07+12dataset reached 86.31%.Of the 20 classes in COCO2014 dataset,m AP reaches 67.71%;The number of network parameters is 10,945,714,and the detection speed reaches 44 FPS.The experimental results show that compared with the basic network(the number of parameters is10,353,068),this network can effectively improve the detection accuracy and speed under the premise of only increasing the number of parameters by a small amount.(2)In order to further reduce the number of parameters and computation of the target detection network and improve the real-time running speed of the network,this paper improved the Ghost Module in Ghost Net and designed a lightweight feature extraction module with adaptive selection of the middle channel.The Module will target information loss function and hyperbolic tangent function combination,according to the current network layer dynamic calculation,select the middle channel containing the most information,reduce the randomness of the channel selection in Ghost Module.The inference speed of target detection can be improved effectively by applying this network to the basic network.The number of parameters is 3,722,192,which is about 60% lower than the base network.The detection speed reaches 59 FPS and the m AP reaches 80.27% on PASCAL VOC07+12 data set.Among the 20 classes in COCO2014 dataset,m AP reaches 62.64%,and the detection accuracy is slightly higher than that of the basic network by 0.15%.The experimental results show that the network can maintain a stable detection accuracy with greatly reduced parameters,and has a certain real-time detection ability,which is suitable for running on mobile devices.(3)In order to make the target detection network out of the laboratory environment and run on the mobile edge device,the network weight file completed by the previous training is converted into om format and transplanted to the Atlas200 developer suite.The network reasoning performance is tested by msame tool.Moreover,local video stream,wired camera video stream and wireless network camera video stream are input into the network to conduct reasoning test.The experimental results show that the object detection network designed and implemented in this paper can meet the real-time video detection requirements of different scenes including multi-target,small target,occluding target and so on.