Research On Aerial Image Target Recogition Algorithm Based On Deep Learning

The wide application of aerial image has affected every aspect of people’s life.The identification and detection of people and cars in aerial image has always been a difficulty in small target identification.UAV aerial image recognition has the following problems.1)Relatively few data sets will increase the cost of target training;2)The small target has fewer feature points,resulting in inaccurate identification;3)The large network structure depends on the computing power of computers and is not suitable for target detection on mobile devices;4)For small occluding targets,the detection accuracy is low and the positioning is inaccurate.In view of the problems mentioned above,this paper makes improvements from three aspects: data processing,feature extraction of backbone network and candidate box regression.First of all,this paper adopts vis Drone2019,an open data set from Tianjin University.Although the number of data sets is large,the scene is single.In order to increase the samples of special scenes in the data set,such as too much light or insufficient light in rainy weather,enhancement processing is made on the data set,including logarithmic transformation,random erasement and Gaussian blur.Secondly,in order to solve the problem of low detection accuracy of small targets,the skeleton network of feature extraction is improved from two aspects of feature fusion and adding receptive field.The feature fusion extractor with efficient-B0 as the skeleton network and the feature extraction network with the combination of RFB module and Darknet53 to increase the sensitivity field were respectively used.The common feature of the two networks is that they are lightweight networks,which can be used for mobile devices,and improve the detection accuracy of small targets by feature fusion and adding receptive fields.Finally,in order to solve the problems of inaccurate positioning of dense small targets and slow detection speed,the positioning loss function was re-selected and the DIOU positioning loss function was used to replace the IOU loss function.In order to deal with the problem of missing small targets,the number and size of candidate boxes were redefined.In combination with Focal loss,the balance of positive and negative samples of candidate boxes was made,which accelerated the detection speed of small targets.