Deep Multi-scale Spatial Adaptive Networks Of Fusion Classification And Object Detection For Remote Sensing Images

High-resolution remote sensing images are often used in urban planning,military and other fields,it is of great significance both in scientific research and practical applications.In this paper,the main research topic is how to solve the problems of fusion classification and object detection for remote sensing images with deep learning.A novel fusion classification network and two effective object detection frameworks are proposed for remote sensing images based on deep convolutional neural networks.1.A novel feature-level fusion classification network is proposed based on for remote sensing images depth adaptive feature learning.This network uses deep learning to extract feature of each image.On the one hand,the information of multi-source remote sensing images are fused on the feature channel,on the other hand,the information of pixel blocks with different size are fused on the feature channel.In addition,the SENet block is followed by each fusion block for the elimination of aliasing,the feature information of different channels have a process of re-fusion.The fusion features are obtained by this fusion structure,which are classified to obtain the results of fusion classification.The key of this network is the combination of multi-source data with different scales of pixel blocks and the addition of the SENet block.The weights of feature channels can be adaptive obtained.The process of adaptive feature learning can satisfy different kinds of land cover,the accuracy of the overall fusion classification network can be improved.2.An object detection framework based on deep adaptive proposal network is proposed for remote sensing images.This framework is based on Faster Region Convolutional Neural Network(Faster RCNN),which is mainly aimed at the phenomenon of uneven distribution of objects in remote sensing images.The category prior network is designed to obtain the priori information,which contains each category number of each image.The fine region proposal network can generates candidate boxes of each image.The adaptive candidate boxes can be generated by the combination of the category number priori information with the candidate boxes.The key of this framework lies in the introduction of the category prior network,which overcomes the problem that the candidate regions is excessive in Faster RCNN,and excessive candidate regions will increase the training difficult of detection networks.Therefore,the adaptive candidate regions are more conducive to the training and testing stage,and further improve the accuracy of detection networks.3.A multi-scale spatial prediction learning network is proposed for remote sensing images object detection,which is based on an improved Feature Pyramid Network(FPN).It is mainly designed for a problem of severely uneven number of positive and negative candidate regions in the deep detection network.We first extract the features by the deep network and obtain the spatial scale information of the object by the spatial scale network.Then the FPN is used for detection,in which the region proposal network will filter most of the candidate regions according to the spatial information of the object,and thus the efficient positive candidate regions can be generated.More importantly,the negative candidate regions can be selected corresponding to the object scale.Finally,each candidate region is used to obtain the final detection results.The core of the proposed framework is that the spatial scale information defines the selection of negative samples,which makes the training stage more directionality and can suppress the complex background information.When compared with the original FPN,the better detection results is obtained by the proposed method.