Research On Small Sample Military Target Recognition Algorithm In Complex Scene

With the continuous development of military science and technology in the new situation,the recognition of military targets with small samples on the ground in complex scenarios is of great significance for future battlefield information joint operations,intelligence acquisition and strategy formulation.Due to the particularity of military targets,it has the characteristics of small samples in complex scenes.In view of the problems of low accuracy,poor real-time performance and weak applicability of existing military target recognition algorithms,this paper takes military targets as the research object and uses deep learning technology.The research is carried out from three aspects: data set construction,recognition model design and model application.The main research contents are as follows:(1)Construction of a small-sample military target data set on the ground in complex scenesIn view of the particularity of military target data,there are very few public military target data sets at present.In this paper,the internal data provided by a research institute of the project cooperation unit and Internet pictures are combined to construct a military target data set.Aiming at the small sample size and sample imbalance of the military target data set,the traditional method and the improved generative adversarial network method are used to expand it.The experimental results show that the obtained images have the characteristics of military targets in complex backgrounds,and the number and size of each category of images are analyzed,and the data set is divided according to the requirements of the military target recognition model for subsequent use.(2)Design of military target recognition model based on improved YOLOv5 algorithmAiming at the problems of high false detection rate and slow speed in the military target recognition algorithm in the complex battlefield environment,this paper proposes the PB-YOLO(Parallel attention Bi FPN-YOLO)military target recognition algorithm.Firstly,the channelspace parallel attention mechanism is introduced to improve the model’s ability to extract small target features in complex battlefield environments;secondly,the Bi FPN weighted feature fusion network is added to improve the model reasoning speed;finally,the Alpha＿IOU loss function is used to solve the target box and prediction box The problem of Io U degradation when overlapping accelerates the convergence of the model and further improves the recognition accuracy;finally,training and testing are carried out on the self-built military target data set,and the recognition accuracy rate reaches 90.17%.Effectiveness of Algorithms in Identifying Military Targets.(3)Lightweight algorithm model design for military applicationsAiming at the limited computing power of embedded devices in military application units,the designed military target recognition model cannot be directly applied in the battlefield.This paper proposes a lightweight YOLO algorithm for military applications.First,the redundant parameters in the PB-YOLO model are deleted through channel pruning.Although this method reduces the size of the model,it cannot improve the model reasoning speed;based on this,the Shuffle Net lightweight module is introduced to reconstruct the backbone network and reduce the amount of model calculation,to improve the calculation speed;finally,the simulation verification is carried out on the self-built data set,and it is concluded that the lightweight YOLO accounts for 26% of the calculation amount of PB-YOLO,and the simplified model size is only 20% of PB-YOLO;secondly It is deployed on an embedded platform for testing,which proves that the designed lightweight YOLO algorithm can have a good recognition effect on military targets in complex scenes on military application unit equipment.