Research And Application Of Hashing Learning Method For Remote Sensing Image

With the development of satellite technology,the number of remote sensing images has increased dramatically.How to better improve the applied research and technological advancement of remote sensing images,has become a hot field in academia and application departments.Due to large data volume,small object size,and complex background of large-scale remote sensing datasets,it is always a key research point for object detection and image retrieval work.For the above-mentioned object detection and image retrieval tasks,traditional methods always suffer from problems such as high computational complexity and memory consume.To address the existing problems,this paper will introduce hashing technology in the research of related theories and algorithms.Overal,by taking advantage of low storage and high efficiency in big data learning and generating a compact hash code,it is feasible to improve the precision and efficiency of remote sensing image object detection and image retrieval.In summary,the main contributions of this paper include:(1)A multi-object classification method for remote sensing images based on affine invariant discrete hash is proposed,which is used to solve the problem of classification perfromance degradation due to the affine transformation of remote sensing image generation.By combining efficient supervised discrete hashing and affine invariant optimization factors,we propose an effective model,namely affine-invariant discrete hashing(AIDH).Specifically,AIDH constrains the affine transformed samples from the same semantic information to similar binary codes to improve the classification accuracy.The experiments on NWPU VHR-10 and RSOD-datasets show that AIDH has higher accuracy and efficiency.Besides,the proposed AIDH provides theoretical basis for object detection in our subsequent task.(2)A fast object detection method based on AIDH is proposed for remote sensing images.Taking advantage of AIDH in multi-object classification,HNM is adopted to generate more effective negative sample set as background class.As a result,the new training set consist of positive and negative samples.Therefore,AIDH can be re-learned to generate a more effective hashing learning machine.By using the selective search results as test samples,the object boxes and the corresponding class labels are detected.Finally,SVM and NMS are used to filter out a few false negative object boxes and redundant object boxes to obtain the final object detection results.Experiments show that the proposed method is more effective to deal with object detection task.(3)An AIDH-CRF model is proposed by combining AIDH with CRF for accurate object detection on remote sensing image.In order to efficiently use spatial information and further improve the accuracy of object detection,we firstly perform the superpixel segmentation on remote sensing image,and then the undirected graph structure is constructed for CRF by taking superpixel blocks as vertices of graph.By using superpixel blocks as test samples,AIDH classification can be considered as unary potential function of CRF to generate the initial category label.Then,the pairwise potential function of CRF is constructed for label re-learning by leveraging Potts model,and the object neighborhood information is smoothed and accuracy object detection is achieved accordingly.Moreover,as a post-processing method,the convex hull boundary is used to generate minimum external rectangular frame to get the final object detection results.The experimental results demonstrates that the proposed method achieves the tradeoff of accuracy and efficiency for objection detection tasks.(4)A low-rank hypergraph hashing(LHH)is proposed for remote sensing image retrieval(RSIR)method.LHH employs l_2-1 norm to constrain the projection matrix to reduce the noise and redundancy among features.In addition,low-rankness is also imposed on the projection matrix to exploit its global structure.Then,LHH uses hypergraph to capture the high-order relationship among data,and is very suitable to explore the complex structure of remote sensing images.Finally,an iterative algorithm is developed to generate high quality hash codes and efficiently solve the proposed optimization problem with theoretical convergence guarantee.Extensive experiments are conducted on three remote sensing image datasets and one natural image dataset that are publicly available.The experimental results demonstrate that the proposed LHH outperforms the other hashing learning in RSIR task.