| As an important part of the computer vision,automatic target recognition has broad applications in the military field.In the actual battlefield,scholars pay more attention on how to accurately identify and mark dangerous targets by quickly analysis and processing massive real-time video data.In this paper,panzers are taken as the main recognition objects.On this basis,the research of panzers recognition technology based on deep learning is carried out.The main work is as follows.Firstly,combined with the latest research of various main algorithms,the YOLO algorithm’s details and network model are optimized.Due to the YOLO algorithm is not sensitive enough to small targets,this paper adjusts the calculation formula for the loss function.In addition,the way of multi-scale training has been added.Considering the missed detection of YOLO algorithm when faces with multi-target clusters,this paper adds five anchor boxes to each grid.In addition,BN layer is added to accelerate the convergence speed of the network.At the same time,residual network is introduced to effectively prevent network degradation.Moreover,1*1 convolution layer is added to extract image details better and improve network recognition accuracy.Secondly,simulating the battlefield environment in a variety of complex contexts and training with simulated data sets.In this paper,an improved marking method for data sets is proposed to avoid overlapping of markup boxes as much as possible.In addition,both YOLO and SSD are trained in this paper,and the algorithms are analyzed and compared in terms of target recognition evaluation indicators.Third,a target recognition system is designed based on the core processing module of NVIDIA Jetson TX2.The YOLO algorithm is transplanted to the embedded development platform,and the target recognition system is applied to the real panzers recognition experiments.The experimental results show that YOLO has strong generalization ability which can meet the real-time requirement of target recognition in battlefield,and perform well in the recognition performance of long-range small targets. |
