Research On Object Detection On The Sea And Land In Aerial Images Based On Deep Learning

Automatic,efficient,and accurate recognition for objects of interest is a key element in the field of aerial images research,and one of the important ways to improve the automation and intelligence of aviation equipment and aircraft.In this paper,we analyze and discuss the key issues of deep learning-based object recognition and detection in aerial images,focusing on the challenges and problems of ship detection on the sea and ground object detection in aerial imagery,and study the role of using virtual engines to build virtual synthetic datasets in view of the lack of aerial datasets and the difficulty of labeling.The main research contents of this paper are as follows:Firstly,18 publicly available datasets for aerial image object detection are collected and organized.The problems and challenges in aerial image object detection are analyzed and summarized in detail with respect to the characteristics of aerial image datasets,mainly including extreme variations in object scale and proportion,smallscale objects,class imbalance,complex environments,dense and occlusion,lack of datasets and difficult labeling,etc.Secondly,to address the problems and challenges of ship object detection,this paper adopts Balanced Feature Pyramids,Guided Anchor and IoU-based sampling methods to improve the original algorithm in the feature fusion stage,region proposals extraction stage,and the positive and negative sample sampling stage of the object detection algorithm,respectively.A large number of benchmark and comparison experiments prove that the three methods effectively improve the detection accuracy of ship object detection in aerial images.The experiments with different types and depths of feature extraction networks also show the effectiveness and consistency of the improved methods in this paper for improving the ship object detection models in aerial images.Again,for the problem of aggregating small object regions in aerial image ground object detection,four data enhancement methods are designed in this paper,including mean crop,mean crop image blended by Mosaic method,extracting dense regions based on Mean Shift method,and dense region Mosaic blend enhancement.Based on the data enhancement methods,the Cascade RCNN algorithm with the HRNet and HRFPN is used for improved aerial image object detection based on high-resolution feature extraction network and feature fusion network;finally,the accuracy and generalization of the object detection model are further improved using the random weight averaging SWA method.According to the experimental results of the paper,the four data enhancement methods and high-resolution networks enhance the detection accuracy of ground objects and improve the detection accuracy of small objects and clustered object areas in particular significantly compared to the non-data enhancement methods.Finally,to address the problems of scarcity of aerial image datasets and difficulties in annotation,this paper completes the study of building a virtual synthetic dataset using Unity virtual engine,and constructs a synthetic ship dataset with a total of 105,086 virtual synthetic images and 194,054 ship objects.All the objects are annotated with horizontal and rotated boxes.The Unity-Ship 10k dataset is used as the additional data to investigate the enhancement of the virtual synthetic dataset to ship object detection in aerial images.The model pre-train and data enhancement experiments were completed respectively.The experimental results show that the pre-training and data enhancement using the virtual synthetic dataset have good enhancement and improvement effects on small and medium-sized objects,while better results are obtained on the single-stage detectors and the Anchor Free detectors.In summary,this paper addresses the problems and challenges in the detection in aerial imagery,and makes corresponding improvements and enhancements in the direction of datasets sources,data enhancement,and object detection algorithms,and proves the effectiveness of the corresponding enhancement and improvement methods through experiments,which effectively improves the accuracy and efficiency.