Research On Cross-domain Object Detection In Remote Sensing Images

Object detection in remote sensing images based on deep learning is highly dependent on data annotation.Algorithms and models trained with single-domain data will consume a lot of labeling costs when faced with various remote sensing detection tasks.Therefore,it is of great significance to utilize easily accessible and annotated datasets to achieve object detection in cross-domain remote sensing images.In this paper,a domain adaptation method is adopted to transfer the knowledge learned by the model in the labeled UAV data to the remote sensing domain,thereby realizing crossdomain object detection on unlabeled remote sensing satellite data.The contributions and innovations of this paper are as follows:（1）In cross-domain detection,the algorithm has limited application and scale limitations of feature alignment.This paper proposes DA YOLO,which designs a plugand-play multi-scale feature transfer module and a more general cross-domain training method.Unlike general cross-domain detection algorithms that can only be applied to two-stage detectors,DA YOLO’s cross-domain detection mode can be applied to various object detection models.Benefiting from the comprehensive feature expression ability of multi-scale features,compared with YOLOv4,the cross-domain detection m AP of DA YOLO is increased by 2.95% to 75.14%.（2）In the process of feature transfer,there are problems that the global features are forced to align and the weights of feature transfer remain unchanged.In this paper,a feature transfer method based on saliency region guidance and a feature transfer method based on weight adaptation are designed.First,in this paper,a visual saliency module is added to DA YOLO to predict salient regions in images,and the prediction results of the visual saliency module are used to guide feature transfer.This ensures that only positive samples of the two-domain image are transferred,and the interference of negative samples such as noise is minimized.Experiments show that the feature transfer method based on saliency regions can effectively improve the cross-domain detection performance of DA YOLO,and the cross-domain m AP is improved by 3.95% to79.09%.Secondly,based on the related principle of H-divergence,this paper takes the classification confidence of the domain classifier as the basis for setting the weights of the two-domain feature transfer.For hard-to-transfer examples,use larger transfer weights,while easy-to-transfer examples use smaller transfer weights.The experimental results show that the weight-adaptive feature transfer method improves the cross-domain detection ability of DA YOLO without adding additional model parameters,and the detection m AP value increases by 1.72% to 80.81%.（3）In order to solve the difficulty of obtaining large-scale labeled data.Aiming at the practical difficulties faced by target detection in the field of remote sensing,this paper constructs a complete closed-loop cross-domain detection process and designs a corresponding semi-automatic labeling tool.Based on the domain adaptation method,the labeling tool designs model training,automatic labeling,manual correction and model training as a closed loop,which greatly improves the labeling efficiency of remote sensing images.Experiments show that compared with the traditional manual labeling method,the semi-automatic labeling method shortens the labeling time by twothirds and greatly improves the labeling efficiency.