Object Detection In Remote Sensing Images Based On Deep Learning

Aiming at the challenges of scale variation of remote sensing images,rotated objects in arbitrary directions,dense arrangement,and a large aspect ratio,this paper improves the generation method of rotated proposal,feature fusion and enhancement,and quality of rotated proposals to improve the performance of remote sensing object detection based on the conventional two-stage object detector Faster R-CNN and single-stage object detector Retina Net.The research in this paper is as follows:1.Object detection in remote sensing images based on multi-scale feature adaptation and rotated box.Based on the objects in remote sensing images are arranged in arbitrary direction,the background of remote sensing images is complex,and the scale of objects varies greatly,this paper proposes a two-stage remote sensing image object detector consisting of a bidirectional fusion feature pyramid,feature adaptive module and rotated region proposal network.We add bottom-up paths into feature pyramid network to enhance the scale-awareness of feature maps at each level.For remote sensing images with complex backgrounds,feature of objects are weak.We use the feature adaptive module to weight the foreground channel to highlight their features.For objects arranged in arbitrary directions,we employ an rotated region proposal network to generate rotated proposals to better adapt to the rotation changes of objects.The accuracy of this method on the DOTA dataset and HRSC2016 dataset is proved to be better than other current remote sensing detection methods through experiments.2.Object detection in remote sensing images based on cascade refinement of proposals and feature alignment.The single-stage detector(Retina Net)proposes a focal loss to better handle the category imbalance problem,and the accuracy of the one-stage detector can be comparable to the two-stage detector,and the model complexity is lower than that of the twostage detector.Therefore,it is widely used in remote sensing images with densely arranged objects.Based on this starting point,we propose an improved one-stage detector,where we utilize a cascaded refinement stage to regress the rotated proposals to make them closer to the ground-truth bounding boxes.To ensure the consistency of features for classification and regression at each refinement stage,we design a rotated anchor constrained convolution.For the regression loss,we design a rotated Intersection over Union loss in terms of Io U to improve the accuracy.Through extensive experiments,the accuracy of this method outperforms other object detection methods on the DOTA and HRSC2016 dataset,which validated the effectiveness of our proposed detector.3.Remote sensing image object detection based on rotated proposal generation network in anchor-free manner.Considering that some object categories in remote sensing images have sharp aspect ratios or are densely arranged,the way to calculate the Intersection over Union between the horizontal anchors and the ground-truth boxes based on the explicitly defined horizontal anchors in the anchor-based detector will result in a lower matching degree or bring redundant background noise problems.To solve this problem,we utilize an anchor-free scheme to adaptively generate the rotated proposals to match the ground truth bounding boxes according to the shape and scale of the objects.After the rotated proposal generation network obtains the initial proposals,we utilize the proposal refinement module to refine the proposals to generate high-quality rotated proposals,where we combine the anchor constrained convolution to realize the classification feature and regression.To further improve the detection performance,we use the improved Fast R-CNN detection head to further classify the highquality proposals and regress the final bounding boxes.Extensive experiments were conducted on the publicly available datasets DOTA and HRSC2016,and our method is proved to be effective compared with the detection algorithms including anchor-based and anchor-free schemes.