| Accurate identification of air vehicle targets is a hot issue in contemporary military field and computer vision field.Accurate and rapid enemy identification,type identification and trend prediction of air vehicles,which provide an important basis for the next decision,are of great importance for the improvement of the efficiency of the future ground fire control command combat system.In recent years,various target detection algorithms developed based on convolutional neural networks have been well used in practice,both in terms of speed to meet real-time and satisfactory results.In this paper,YOLO-v3 target detection network is introduced to study the problem of accurate recognition based on aerial vehicle images,and the main work is as follows.1)To address the problem of insufficient cross-scale and cross-space feature fusion,this paper optimizes the feature fusion method of YOLO-v3.The feature pyramid transformer FPT is adopted as the feature fusion method of detection network,which combines the advantages of Non-Local and FPN,and adopts different operation methods to fuse the features on cross-space and cross-scale through the design of three transformers,thus realizing the effective use of contextual information and information on different scales,and to a certain extent,improving the the detection performance of YOLO-v3,and,in terms of temporal performance,does not differ significantly from the original network.2)To address the problems of inaccurate selection of overlapping frames of the intersection and ratio functions and the imbalance between positive and negative samples of air vehicle images,the loss function used in YOLO-v3 is improved in this paper.Considering the small target and complex background of air vehicle image recognition,three types of losses need to be considered for the loss function of the model:coordinate loss,classification loss and confidence loss.The generalized intersection ratio function can have a certain correction effect on the border loss,while the Focal Loss can reduce the weight of the model on the hard-to-score samples in the sample set and help the model to be able to learn more effective features,thus alleviating the confidence and classification losses.3)Experimental results and analysis.In this paper,experiments are set up to verify the effectiveness of the prior frame selected for clustering,the FPT fusion method and the optimized loss function,respectively.The aerial vehicle dataset was constructed with comprehensive consideration of scene and illumination,and the labeling process was performed to obtain the training and testing sets required for the network model to perform learning.In order to improve the convergence speed of the network,the K-mean clustering algorithm was used to select the prior frame before the network training.For network training,DarkNet-53 was pre-trained on the COCO dataset using the idea of migration learning as a feature extractor.The target detection network containing FPT is used to detect air vehicles,and the average accuracy is improved by about 1.6 percentage points compared with the original network.Based on this,the average accuracy can reach 63.3%using generalized cross-merge ratio and Focal Loss loss function,which is about 2.3 percentage points better than the original network.The experimental validation results show that the optimization scheme proposed in this paper based on YOLO-v3,the feature fusion method and the loss function is feasible and effective for aerial vehicle image recognition,and has certain. |
