Object Detection In Remote Sensing Images Based On Multi-scale And Multi-oreintation Features

Recently,object detection in remote sensing images plays an important role in disaster monitoring,urban planning and other fields.Compared with natural images,scale and orientation of objects in remote sensing images change a lot due to the different perspective and imaging characteristics.Therefore,objects in remote sensing images have more salient multi-scale and multi-directional features,which makes it hard for accurate classification and localization,and significantly affects the detection performance,leading to great challenge to object detection in remote sensing images.To handle the above issue,object detection in remote sensing images is studied on the basis of multi-scale and multi-orientation features.Firstly,to address the multi-scale problem of remote sensing images,adaptive multi-scale region of interest pooling(AMRP)is designed,which is composed of the generation of multi-scale region of interest pooling(Ro I pooling)and attention mechanism based multi-scale feature fusion.AMRP can not only provide multi-scale features,but also retain the function of Ro I pooling that mapping any size of features to a fixed size.Then,rotated feature network(RFN)is proposed to solve the multi-orientation issue,which consists of Encode and Decoder units.Different from previous works,the classification and regression tasks do not share the same features.RFN provides rotation-invariant and rotation-sensitive feature maps for the two tasks respectively,so that the conflict problem can be handled that classification task is not sensitive to orientation while regression task is quite sensitive.Moreover,to ensure the reliability of the rotation-invariant characteristics,rotation-invariant loss is proposed in the training phrase.In this thesis,experimental analysis is used to verify the effectiveness of the proposed methods.The experimental results of multi-scale object detection in remote sensing images show that AMRP method can improve the performance of Faster R-CNN by 2.3%,and also significantly decrease its training time,accelerating the convergence of neural network.In addition,compared with several popular methods,AMRP can improve the detection performance by at least 4.7%.The experimental results of multi-orientation object detection in remote sensing images show that RFN method can improve the performance of Faster R-CNN by at least 5.4%,and also outperforms several currently popular object detection methods.RFN is further evaluated on scene classification of remote sensing images and object detection of natural images,demonstrating good generalization performance.RFN can be integrated into the existing frameworks,leading to great performance with a slight increase in model complexity.