Research On Aerial Target Recognition Method Based On Deep Learning

In modern warfare,air combat weapons play an increasingly important role,even affecting the overall trend of the situation.By identifying aerial targets captured by image acquisition equipment,we can use the acquired information to achieve effective friend-or-foe identification.This information can also be analyzed to determine the source of the enemy,combat capabilities,combat intentions,etc.,and provide an important reference for our tactical decision.Therefore,aerial target recognition in actual combat scenarios has become an important part of the military offensive and defense systems of various countries.With the continuous development of military engineering technology in various countries,various types of air combat weapons have gradually derived a large amount of models based on different operational purposes and functions.In addition,with further implementation of the concept of information-based warfare in our country,the channels for obtaining air battlefield information are expanding and information obtained is increasing.Therefore,the traditional artificial recognition methods can no longer meet the task of aerial target recognition under the current situation.In recent years,deep learning plays an increasingly prominent role in the field of target recognition,and has gradually replaced the traditional image processing algorithms.In this paper,the aerial target recognition task is decomposed into two processes: aerial target detection and aerial target type recognition.The key technical problems that need to be solved in each process are summarized firstly,the deep learning-based methods to carry out the research.The specific research contents are as follows:(1)About aerial target detection algorithm,this paper studies the existing deep learning target detection algorithms,chooses Faster R-CNN target detection algorithm as the research basis after analysis and comparison,and proposes two improvements based on Faster R-CNN to better adapt to the task of aerial target detection.First,feature pyramids are generated based on the backbone network of the original Faster R-CNN algorithm,and make full use of the feature maps of different levels to effectively extract the features of small-scale targets.On the other hand,an improved Anchor matching rule is proposed.This matching rule can make all ground truth boxes match more Anchors than the original matching rules during the Anchor matching process.(2)About aerial target type recognition,to solve the problem that the target detection algorithm can only classify the aerial target roughly,this paper designs an aerial target classification network based on deep neural network.This paper studies the network structure of the commonly used deep learning classification network,and analyzes their advantages and disadvantages.Finally,the Dense Net is chosen and further optimized to be more suitable for aerial target model classification tasks.(3)In order to solve the problem of lacking suitable aerial target dataset in military field,this paper labels and constructs an air dataset.This dataset contains 15,000 images and a total of 21,726 targets,including6 categories of aerial targets: fighters,bombers,Military Unmanned Aerial vehicles,transportation aircraft,early warning aircraft and civil aircraft,and involving 57 specific models,such as F-22,Su-27,E-3,etc.In this paper,this dataset is used to train and test the aerial target recognition algorithm.Based on the test results on the self-built aerial target dataset,this paper verifies that the target detection algorithm designed in this paper can effectively improve the multi-scale aerial target detection accuracy,especially small-scale aerial target detection accuracy,and also verify the validity of the aerial target identification network designed in this paper.(4)In order to deploy the aerial target recognition algorithm in practice,this paper designs and develops a prototype aerial target recognition system based on C/S framework,which is used to deploy the proposed aerial target recognition algorithm based on deep learning in this paper;tests this system,and the test achieves relatively satisfactory results.