Aerial Image Object Detection Based On Deep Learning

With the development of drone technology and satellite remote sensing technology,a number of very representative products have emerged,such as: DJI drones,and satellite photos,web platform which combined aerial photographs and geographic information system photos.The aerial image data obtained through the above technologies has also become richer and more diverse,and its acquisition channels are more convenient and open.At the same time,the research and application of deep learning in the field of computer vision continue to make progress,providing excellent processing methods such as semantic segmentation,object detection,etc.Therefore,using computers instead of humans to process and analyze aerial image data has become a research hotspot.The task of object detection on aerial images is frequently used in application scenarios such as drone detection,traffic supervision,land surveying,wildlife tracking,and disaster observation.It is the basis for many subsequent studies.Based on the current research status of image object detection,this thesis selects a one-step target method suitable for real-time processing of devices such as drones and monitors,and applies the YOLO model to object detection of aerial images.The main content includes two parts: pre-processing and object detection.Pre-processing usually refers to enhancing the image,including operations such as rotation change and defogging,expanding data samples for subsequent object detection tasks,and providing high-quality image data.The object detection of aerial images uses the YOLOv3 model based on Darknet53 for experiments.It borrows common excellent models such as residual networks and feature pyramid networks to improve accuracy.Experiments are performed using public data sets,and the results are analyzed.Then,based on the characteristics of aerial images such as large image resolution but small target objects,improvements to the network model are proposed.Based on the principle analysis of YOLOv3,kmeans clustering was carried out to obtain the anchor box size of the aerial image data set.As the shallow feature maps often have a smaller perceived field of view,the deep feature maps have a larger perceived field of view,which is important for prediction of large object size differences in aerial images.Therefore,try to increase the resolution of the input picture and combine Deeper residual network for feature extraction,then using the idea of feature pyramid network,using feature fusion to enhance the feature map,increase the information contained in the feature map participating in the prediction,through experiments,it can improve the prediction of objects with large scale differences effect.After that,the thesis tried to improve from the loss function and compared the experimental results.The thesis discussed possible causes of this phenomenon that the results found that the prediction effect did not improve much.Finally,the improved model is compared with the existing target detection methods,and it is found that this method is more suitable for real-time target detection of aerial images.