Research On Aerial Vehicle Detection Algorithm Based On Deep Learning

The in-depth development of artificial intelligence technology has greatly improved the efficiency of life and production,and effectively promoted the intelligent process in the fields of industry,agriculture,transportation,medical treatment and so on.The technologies represented by computer vision,speech recognition and natural language processing have broad prospects for industrial development.Based on this background,this topic is carried out from two aspects of deep learning and computer vision,focusing on the research of vehicle detection algorithm in aerial scene,so as to improve the detection effect of aerial vehicle in natural scene.Firstly,this paper describes the development status of object detection technology at home and abroad,summarizes the relevant detection algorithms and research,and summarizes the relevant research methods and results.Then it analyzes the basic process of traditional object detection algorithm and depth learning algorithm,and introduces the aerial image detection data Vis Drone.Finally,the object detection model and the design idea of detection framework are analyzed and summarized.Due to the small object size in the aerial scene,this paper proposes an aerial vehicle detection method based on feature recursion to improve the detection effect of small objects.Based on the two-stage detection algorithm Cascade R-CNN,this method designs the detection model based on Feature-based multi-path recursive fusion.Firstly,based on the classical feature pyramid network,using 3×3 after adjusting the number of channels,the features of the output stage are recursive to the top-down path for feature map fusion at the same level.At the same time,using the sampling method CARAFE based on deep learning,the feature map of the output stage is up sampled to generate a new feature map,and the feature map is recursive to the next layer of the previous stage structure for fusion;The fusion structure effectively couples the features of different scales and reduces the semantic gap between features.Experiments show that compared with the improved feature pyramid network,the detection accuracy of small objects and medium-sized objects based on multi-path recursive fusion method is improved by 1.0% and 1.6% respectively.Based on the recursive feature fusion structure,a multi-scale atrous convolution detection algorithm is proposed to improve the problem of large size change in aerial scene.By designing convolution kernels with different atrous rates,the algorithm constructs a feature fusion structure of multi-scale atrous convolution,which effectively improves the effect of information capture and feature reorganization in the feature pyramid network in object detection.The improved algorithm adopts the characteristics of C2,C3,C4 and C5 output in the residual network,and carries out atrous convolution operations with convolution kernel size of 1,3 and 5 and expansion rate of 2,1 and 3 for each level of characteristic map respectively.The convolution combination of multi-scale atrous rate has different receptive fields,the structure effectively alleviates the gridding effect caused by increasing the dilated rate,and the optimized algorithm improves the accuracy of large object detection by 3.9%.