Research On Remote Sensing Image Retrieval Based On Deep Metric And Category-level Semantic Hashing

With the rapid development of remote sensing satellite technology,we have entered the era of remote sensing big data.The storage and retrieval of high-dimensional data has become a bottleneck in the application of massive remote sensing data.How to achieve accurate and fast search of large-scale remote sensing images is of great significance in remote sensing applications such as disaster warning,agricultural detection,and urban planning.As an efficient indexing method,the hash method generates a short binary sequence representation for input image,which greatly improves retrieval efficiency and reduces memory consumption simultaneously.However,the traditional hash methods rely on artificially designed feature descriptors as the input of hash coding,and the artificially extracted low-level visual features are difficult to describe the complex visual information of remote sensing images,which restricts the improvement of hash retrieval performance.With the development of deep learning,deep convolutional neural networks can automatically extract more abstract deeplevel features of images,and the learned features have stronger discriminative capabilities.Therefore,people propose a new method that integrates deep learning and hashing learning,called deep hashing,which has achieved satisfying reteieval performance.However,due to the complex backgrounds and structures of remote sensing images,the research on deep hash retrieval of remote sensing images with semantic similarity preservation still remains a challenging issue in remote sensing image retrieval.Main contributions of this thesis are:Aiming at the loss of feature information in the hash mapping process,in order to generate similarity-preserved hash codes,this thesis proposes a simple and effective deep metric and category-level semantic hash network model named DMCH.The model is mainly divided into feature extraction and hash codes.First,the Inception Net network pre-trained on the Image Net dataset is used as the feature extractor to extract the high-level semantic information of the remote sensing images as the intermediate representation,which greatly reduces the training time and also makes up for the lack of the number of labeled samples in the remote sensing image dataset.In the hash functions learning stage,the proposed model selects effective triplets to learn the hamming metric space,adds category-level classification loss and ideal hash code constraints to enhance the potential relevance between hash codes,and generate concise,effective,and more discriminative hash codes,thereby to improve retrieval performance.Aiming at the problem of using hamming distance to measure the similarity of hash codes with large granularity,this thesis introduces two rearrangement algorithms to fine-grained reordering the retrieval results of the model DMCH,which are based on the weighted hamming distance rearrangement algorithm and the hash rearrangement algorithm from coarse to fine.The weighted hamming distance rearrangement algorithm based on the important bits of the nearest neighbor distinguishes the relative importance of the bits according to the difference distribution of the hash code in each bit in the retrieved images,increases the weight of primary bits and decreases the weight of secondary bits.The coarse-to-fine hash rearrangement algorithm recorders the similarities by comparing the more detailed real-valued features vectors before relaxation,and is more suitable for distinguishing the similarities between images with the same hash codes.Finally,this thesis verifies the effectiveness of the proposed model on three public standard optical remote sensing datasets through a series of comparative experiments.The experimental results prove that the deep metric and category-level semantic hash network model proposed in this thesis are effective in remote sensing image hash retrieval performance.Specifically,compared with the current more advanced hash methods,the retrieval accuracy of the model on the three datasets has been improved by 1.13%,2.89% and3.36%.In addition,the two rearrangement algorithms have further improved by 0.57%,0.2%,1.28%,and 0.98%,0.6%,2.35% by fine-grained optimization of the retrieval results of the model.Among them,the coarse-to-fine hash rearrangement algorithm can improve the retrieval performance of the proposed model more obviously.In summary,this thesis proposes a deep metric and category-level semantic hash network model and introduces two hash rearrangement algorithms,which have some theoretical and application value for research on deep hashing with semantic similarity preservation for remote sensing image retrieval.