Research On Object Detection Algorithm Of Remote Sensing Images Based On Deep Learning

With the advancement of remote sensing technology,the amount of remote sensing data acquired by people is rapidly increasing,so that more efficient and automated means of processing data are needed.As the most popular artificial intelligence technology,deep learning has become a hot research field,and has shown potential in the field of remote sensing image object detection.This paper focuses on the remote sensing image object detection algorithm based on deep learning.This paper analyzes the shortcomings of existing algorithms,and conducts further research in combination with the task of remote sensing image object detection.First of all,the objects in the remote sensing image are of different shapes and sizes.And there are many small objects among them.These small objects have less information,are not easy to identify,and are easily lost in the network.This makes the detection effect of the network on small objects poor,and the situation of missed detection and wrong detection is serious.Secondly,as the resolution of remote sensing images becomes higher and the performance of algorithms improves,the demand for high-quality detection based on higher IoU thresholds also increases.However,objects with different shapes and sizes in remote sensing images are a difficult problem for the detector.These objects cannot be accurately positioned at the same time to achieve high-quality detection,and in most cases,the IoU threshold standard for detection can only be reduced.In response to the above problems,this article carried out the following work:(1)Introduce feature pyramids and use structured anchor design.Suppress the selection of anchor points on top-level features,and enhance the selection of anchor points on bottom-level features.With the enhanced method of anchor point selection,the anchor points containing small objects are given more training opportunities to guide the network to detect small objects.(2)Use a multi-stage cascade detector to improve the quality of the candidateframe layer by layer,so as to approach the real reference frame to achieve high-quality detection.At the same time,the frame regression function is improved,so that the candidate frame tends to be more stable during the regression process,especially when the candidate frame is close to the real reference frame,making the detector more stable at a high threshold,thereby improving the accuracy of high-quality detection.Through experimental verification,the method in this paper has greatly improved the recall rate of small target detection problems,and the situation of wrong detection and missed detection has been improved.At the same time,the method of this paper is more accurate for the positioning of objects,which greatly improves the detection accuracy under the high IoU threshold.