Object Detection In High Resolution Remote Sensing Imagery Based On Deep Hashing

Object detection is an important branch of remote sensing application technologies.In recent years,it has played increasingly important role in many military and civilian fields.Recently,deep learning has been used for automatic object feature learning,which overcomes the problems of insufficient feature extraction and low automation of traditional algorithms.However,with the development of remote sensing imaging technology,both the image size and image complexity increase rapidly,which poses a challenge for fast object detection in wide-range scenes.Hashing learning has widely used in image retrieval,which maps data into binary code through machine learning algorithms and,effectively improves computational efficiency with its binary calculation property.In this paper,we introduce hash learning into the two-stage deep learning algorithm for object detection of highresolution and wide-range remote sensing images.In our work,the hashing algorithm is combined with deep learning from three aspects: the structure,feature and mechanism to develop three deep hashing algorithms for fast object detection,to achieve high detection accuracy while improving the detection efficiency.The main work and innovations are as follows:1)Aiming at the problem of non-dense distribution of objects in wide-range remote sensing images,a deep hashing aided model is proposed in this chapter.Firstly,a hash aided branch is constructed in the two-stage detection model to quickly determine whether the local patch is a potential target area,then we construct a cooperative hash regular loss,and perform endto-end training on the multi-branch network.On the one hand,the hash aided branch has the characteristics of simple structure and efficient binary calculation,and the additional calculation introduced by the discriminant process can be neglected;on the other hand,the hash aided branch can avoid a large number of invalid calculations of the detection model on the non-target area to save detection time.We perform analysis and verification on several datasets such as IPIU,NWPU VHR-10.Experiments show that the deep hash aided network can reduce the computation time on non-target areas by 50.91% to 59.40%,and the detection result MAP is also improved by 0.48% to 1.99%,which verifies the effectiveness and efficiency of the network.2)In order to further improve the speed of object detection,a deep discriminative hash network is proposed.Firstly,the model head parameters with high computational complexity in the Faster-RCNN model are simplified,and the feature dimensions are reduced,then the L1 quantization loss is designed to obtain the binary hash feature.In addition,for the precision loss caused by parameter reduction,a discriminatively guided hash loss function is designed to train deep discriminative hash network is designed.We perform analysis and verification on several datasets such as IPIU,NWPU VHR-10.Experiments show that compared with the baseline model Faster-RCNN,the detection speed of this algorithm is increased by 32.85% to 53.93% on different datasets,the detection accuracy also has an increase of 0.39% to 1.55%,which proves the efficiency and accuracy of the method.3)Aiming at the problem of insufficient local semantic information in existing detection models,a deep multi-attention hash network is proposed.The attention mechanism is introduced in the deep discriminative hash network,and the spatial attention and channel attention modules are constructed respectively.On the one hand,the spatial attention module is used to introduce the context information to enhance the semantic representation of the local features,and at the same time suppress the complex background interference in the remote sensing images.On the other hand,the channel attention module adaptively enhances the beneficial features and suppresses the useless features to improves the detection accuracy.We perform analysis and verification on several datasets such as IPIU,NWPU VHR-10.Experiments show that compared with the deep discriminant hash network,the detection result MAP of this algorithm is improved from 1.50% to 3.33%.