Research On Few-shot Object Detection Algorithm For Optical Remote Sensing Satellite Image Interpretation

As computer vision technology continues to advance and optical remote sensing technology rapidly develops,the application of optical remote sensing images is becoming more prevalent.Object detection tasks on optical remote sensing images have become more and more important in military,environmental monitoring,urban planning,land use and other fields.Research on object detection methods for optical remote sensing satellite images is of great significance for improving the automation level of remote sensing image processing,promoting the development of related fields,and has important implications.The targets in optical remote sensing images have many characteristics,such as significant aspect ratios,arbitrary orientations,generally smaller scales,and complex background information,which make the interpretation of these images challenging for conventional object detection algorithms.Currently existing remote sensing image object detection algorithms still have considerable room for improvement.Furthermore,the difficulty in acquiring optical remote sensing images,data annotation,and low data quality further limit the performance of optical remote sensing image object detection algorithms.Therefore,this paper will focus on the research of object detection on remote sensing images and few-shot object detection.Addressing these challenges will provide more accurate and efficient methods for the automation of optical remote sensing image processing,promoting the further development of related fields.In terms of remote sensing image object detection,this paper proposes a new rotated cascade region proposal network(RCRPN).Based on the two-stage detector Faster RCNN framework,RCRPN is an improvement to the RPN network.The RCRPN module is divided into two stages.The first stage uses the design of a single anchor box at each position on the feature map and uses the midpoint offset representation method to encode the rotated bounding box,thus avoiding direct regression of the angle of rotated bounding box and outputting coarse-grained proposals.The second stage takes the proposals generated by the first stage as input,and uses adaptive alignment convolution to extract features and refine the proposals,thereby obtaining high-quality rotated proposals.Compared with the baseline method,RCRPN improves detection performance while reducing memory consumption,providing a superior basic object detection model for downstream tasks.Regarding few-shot object detection,based on the work of RCRPN,this paper proposes a remote sensing object detection algorithm under a few-shot setting.This study first improved the backbone network and then used domain adaptation-based transfer learning to train the network,thereby overcoming the problem of model overfitting under few-shot conditions.With this method,the algorithm showed excellent detection performance in 2-shot,5-shot and 10-shot experiments on few-shot datasets,demonstrating good generalization ability.To verify the detection performance and robustness of the proposed algorithm,this study collaborated with Chang Guang Satellite Company.to use Jilin-1 satellite image data to construct a small dataset for few-shot object detection on remote sensing images,including various fine-grained annotations for ships and airplanes,which simulated the real optical remote sensing satellite image interpretation scenarios to test the algorithm’s performance.Experimental results showed that the remote sensing image object detection algorithm proposed in this paper had high detection performance and robustness under both fully supervised and few-shot settings.In addition,this study also conducted application verification on Synthetic Aperture Radar(SAR)image datasets such as SRSDD and HRSID,proving that the algorithm also has good performance in SAR images.