Research On Remote Sensing Image Object Detection Method Based On Deep Learning

With deep learning methods gradually replacing traditional manual feature extraction,object detection,one of the most dominant tasks in computer vision,has made significant progress in recent years.A series of excellent object detection algorithms based on convolutional neural networks have emerged in natural scene images.However,there are significant differences between remote sensing images and natural scene images.Remote sensing images are mainly captured by satellites at a top view angle and the background is more complex,thus introducing a series of special problems such as large inter-object scale variations,many small and dense objects,arbitrary rotational orientation of targets,and scarcity of labeled samples,which pose a greater challenge to object detection.To address the above problems,this paper takes the detection of undirected and directed remote sensing images as the main starting point,and proposes the corresponding solutions respectively.To address the problems of sparse samples and small and dense objects in undirected remote sensing images,this paper proposes a novel small sample detector with contextual information refinement.First,a context information refinement(CIR)module is designed to extract discriminative contextual features,mainly by capturing contextual information of different receptive fields through dilated convolutions and dense connections.In addition,this paper improves the region proposal network(RPN)by fine-tuning it on the new category and softening the non-max suppression(NMS)constraint,which can obtain more positive anchors for the novel category.This method achieves 5%-20%improvement over the baseline on two undirected remote sensing public datasets.To address the problem of objects with arbitrary rotational orientation in directional remote sensing images,this paper designs a Transformer-based two-stage directed remote sensing object detection algorithm,which generates directional proposals in the first stage and performs directional bounding box regression and classification in the second stage.In this paper,a oriented region proposal network(oriented RPN)is used to generate high-quality rotational proposals and extract rotationally invariant features at a light cost.In addition,this paper is the first application of Swin Transformer to directed remote sensing detection.Swin Transformer improves computational efficiency by performing self-attentive computation in nonoverlapping windows and allowing cross-window connectivity,enabling the extraction of richer global contextual information.This method achieves 2%-8%improvement over the baseline on two oriented remote sensing public datasets.