Research On Lightweight Technology Of Multi-Target Recognition Deep Learning Network

Multi-target recognition is one of the core functions of airborne target awareness system,which needs to be highly integrated with the airborne platform system.Deep learning target recognition technology provided a new way for airborne target perception system and has broad application prospects.However,in porting deep learning networks to hardware platforms,not only model accuracy but also the number of network parameters,memory resources of aircraft hardware platform and arithmetic power need to be considered.When deployed on an airborne hardware platform with limited resources,adapting the deep learning network to the airborne hardware platform is a technical challenge that must be solved when a large scale,computationally complex multi-target recognition deep learning network was applied.This paper relies on the project of the Aeronautical Science Foundation to carry out research on the lightweight technology of deep learning network,which is aimed at the balanced designing between the target recognition performance and the resource elements required for algorithm operation,improving the adaptation degree between deep learning network and embedded hardware platform.The main research contents are as follows:1)Deep learning network structure analysis and multi-target recognition image database construction were done.Works included to analyze the composition of deep learning network and the working principle,to study the current mainstream deep learning network lightweighting and acceleration methods,to complete the construction of multi-target image database according to the project index requirements,to use the data enhancement algorithms to enhance the image data,sufficiently ensuring the authenticity and diversity of multi-target recognition image data.2)A YOLOv5s network loss function optimization and global channel pruning network lightweighting algorithm were studied.In order to improve the accuracy of the Algorithm’s airto-ground observation multi-target recognition,the loss function of the original YOLOv5s deep learning network was improved for using the CIOU loss function,so as to enhance the overall recognition accuracy of the algorithm.To address the problems of redundant parameters and complex computation of the target recognition network,a global channel pruning algorithm was proposed to train sparse scaling factor γ coefficients on the BN layer of the deep learning network,screen out the parameter transmission channels that can be optimized in the network,and use the global channel pruning algorithm to prune the deep learning network while ensuring the recognition accuracy of the network to generate a lightweight multi-target recognition network with good detection performance.3)A compact network architecture design and research with embedded implementation technology were done.It was found that the depth-separable convolution was significantly less computationally intensive than the standard convolution in the extraction of target features.Therefore,the CSPDarkNet53 backbone feature extraction network using standard convolution in the YOLOv5 model with optimized loss function was improved to MobileNetv3 lightweight network structure using the depth-separable convolution,constructing a compact network YOLOv5-C-Mv3.Not only can the feature extraction ability of the network model be enhanced,but also the number of parameters and the amount of operations are significantly reduced,which provides a new idea for the problem of lightweighting deep learning networks.Based on the software architecture of the NVIDIA Jetson TX2 embedded image processing platform,the deep learning development environment was configured to deploy and compile the lightweight target recognition network on the embedded image processing platform.The experiments show that the multi-target recognition deep learning network lightweighting method studied in this paper can effectively reduce the number of parameters and operations of the deep learning network with small variations in target recognition performance,and facilitate the deployment of the target recognition network on resource limited airborne embedded hardware devices.