The Research On Small Object Detection Algorithm In Aerial Images Based On Depth Neural Network

Object detection is a technology for locating and classifying objects in images.It has important scientific significance and application value in the field of image processing.This paper aims at the requirement of real-time small object detection in high resolution aerial image.Through analysis of the difficulties of object detection in aerial images and comparing the detection accuracy and speed of state-of-the-art algorithms,this paper chooses YOLOv3 algorithm as the basis for small object detection in complex background.YOLOv3 algorithm has superior performance,but it has the problems of small object miss detection and false detection under background interference in aerial image detection.Based on in-depth study of YOLOv3 algorithm and aerial image characteristics,this paper improves YOLOv3 algorithm,improves the performance of small object detection in aerial images,and achieves real-time detection speed.In this paper,the major contents and innovation points can be summarized as follows:(1)Aiming at the requirement of model training of small object detection in aerial images,this paper makes a small object data set of aerial images based on five public aerial remote sensing image data sets.The object categories of the data set are airplane,car and ship.To solve the problem that the initial anchor boxes size of YOLOv3 algorithm does not match the objects size of aerial images,which leads to the decrease of detection accuracy,the initial anchor boxes are redesigned according to the scale distribution characteristics of aerial objects.This paper uses the K-means++ algorithm to cluster the objects in aerial images and uses the clustering results as the initial anchor boxes.Through the test of aerial small object image,the detection m AP using clustering anchor boxes is 0.5% higher than the original algorithm,and it will not affect the detection speed.In this paper,clustering anchor boxes is used in the subsequent improved algorithm.(2)To solve the problem that small-scale objects are easy to miss detection in aerial images,this paper presents an improved YOLOv3 algorithm with high resolution feature fusion,named YOLOv3-F.Compared with ordinary images,the proportion of small objects in aerial images is high and the object scale is small,while the resolution of YOLOv3 algorithm detection layers is low,which can not effectively describe the feature of small objects,and is not suitable for the accurate detection of small objects in aerial images.This paper fuses a higher resolution feature layer based on the original detection layer of YOLOv3,which effectively improves the detection accuracy of small objects.When the input image resolution is 416?416,the detection m AP of YOLOv3-F algorithm on ASOD data set is 89.1%,which is 1.6% higher than YOLOv3.Moreover,the detection speed of YOLOv3-F algorithm on GTX 1060 graphics card is 21.9 FPS,which is only 9.9% lower than YOLOv3.(3)In order to solve the problem that objects under background interference are easy to false detection in aerial images,this paper proposes a new receptive field enhancement structure,named Dense RFB,which combines the receptive field block of RFBNet and the dense connection idea of Dense Net,then designs an improved algorithm YOLOv3-DF embedded in Dense RFB structure based on YOLOv3-F,which can further improve the detection accuracy and the detection speed loss is very small.When the input image size is 416?416,the detection m AP of YOLOv3-DF algorithm on ASOD data set is 90.4%,which is 1.3% higher than YOLOv3-F and 2.9% higher than YOLOv3.Moreover,the detection speed of this algorithm on GTX 1060 graphics card is 20.8 FPS,which can achieve real-time detection.In order to verify the generalization ability of Dense RFB structure,this paper tests the performance of the improved algorithm embedded in this structure on VOC2007 data set,the detection m AP reaches 80.8%,which is 3% higher than YOLOv3.