Research On Aerial Image Object Detection Algorithm Based On Convolutional Neural Network

Unmanned aerial vehicles(UAVs)are widely used in military,disaster relief and daily life because of its convenience,low cost and high efficiency.With the rapid development of UAV technology,a large number of aerial images which contain abundant information and have important research significance and application value have been generated by UAV.In recent years,the development of deep learning technology has significantly improved the accuracy of object detection in natural images.However,we can not get satisfactory results when these object detection algorithms are directly applied in aerial images without modification.This thesis mainly studies on the task of aerial image object detection.Aiming at the difficulties and problems in aerial image object detection,two improved methods are proposed.The main work of this thesis is as follows,In this thesis,a difficult regions re-detection algorithm in aerial images is proposed to solve the problems of uneven distribution of objects and large difference in object size.In order to solve the common problem of uneven distribution of objects in aerial images,we use the clustering algorithm based on density to search the clustering regions of dense objects on the coarse detection results of aerial images.Then the clustering regions are divided into simple regions and difficult regions by calculating difficulty scores.Simple regions only need coarse detection,while difficult regions need fine detection to obtain higher detection accuracy.Because of the large scale differences in difficult regions,the scale differences among objects are further enlarged.In order to solve this problem,Gaussian scaling function is proposed.The Gaussian scaling function can calculate the scaling factor for each difficult region to reduce the scale difference of the object.Finally,the fine detection of the difficult region and the coarse detection results of the image are substituted into the Soft-NMS algorithm to get the final detection results.The experimental results show that the difficult regions redetection algorithm in aerial image on Vis Drone dataset has better performance than other re-detection algorithms,especially in improving the detection performance of small objects.In order to solve the problem of class imbalance in aerial images,an adaptive resampling technique for object of minority classes is proposed in this thesis.Resampling techniques balance the amount of information between minority and majority classes by copying and pasting object of minority classes.Random pasting of resampled objects into an image will result in background mismatches,size mismatches,and brightness mismatches.In view of the above three kinds of mismatch problems,this thesis puts forward the solution.In this thesis,the segmentation network is used to segment the road and find the appropriate paste position for the resampled object to solve the problem of background mismatch.A linear regression function is used to estimate the appropriate size for resampled object to solve the problem of scale mismatch.Then,the resampled object is converted from the RGB space to the HSV color space,and the brightness of the resampled object is adjusted to the brightness similar to the image through linear transformation to solve the brightness mismatch problem.The object after scaling and brightness adjustment is pasted to the selected resampled position to complete the resampling of the object.The experimental results show that the network trained by the enhanced data based on adaptive resampling technique for object of minority classes has higher detection accuracy,especially improved detection accuracy of the minority classes.