Research On Remote Sensing Image Retrieval Method Based Deep Hashing

Remote sensing image retrieval is aiming to obtain images,which are similar to an inquiry image,from a large number of images.Different from the image in the natural scene,the remote sensing image is usually the top view,and carries more band information,so it requires higher feature dimensions to express,which leads to low retrieval efficiency.Recently,deep hash methods have presented promising performance in reducing dimensions of features.However,compared with the natural image,the remote sensing image has larger similarities,correlations and smaller interclass distances,which hinder the application of deep hashing methods in remote sensing image retrieval.In order to solve these problems,improve the quality of hash codes and make retrieval more accurate,this paper develops the remote sensing image retrieval method based on deep hashing method,mainly focusing on the problems of improving differences between images in different classes and enlarging interclass distances.The main works are as following:Firstly,for overcoming the problem which has greatly similar features between different classes in remote sensing images,this paper combines the deep hashing method of point labels with the block connection layer,and proposes the slice-feature based deep hashing method of point labels.This method maintains the semantic information and enlarges feature differences among different classes.In order to extract effective features of images,this method fine-tunes the pre-trained network model,makes model fit features of remote sensing images.In the meantime,it ensures the accuracy of features and the reliability of hash learning and promises the retrieval results to be similar to the inquiry image.For reducing the problem of large similarities between different classes,this paper proposes the feature block strategy to improve the fully connection layer for features,which enlarges differences of hash codes in different classes.Because the number of available remote sensing images is limited,this method adds a regular term to the objective function,improves the robustness of the model,and avoids the phenomenon of over fitting.The experimental results show that the method enlarges the differences of features and maintains the semantic features.Then,when there are more similar classes in remote sensing image,the deep hash method based point labels cannot distinguish differences of features,so we require to further enlarge interclass distances.In order to achieve this goal,this paper proposes a slice-feature based deep hashing method of triple labels to complete metric learning and enlarge interclass distances.In this method,the pre-trained network model is employed as a feature extractor,and a small triplet network is designed for hash learning.The slice layer is used to reduce dimensions for improving the effectiveness of features and promising the training effect of the model.In order to further enhance quality and correction of hash codes,this method adds loss terms in the loss function of network model.Based on keeping the feature differences,this method realizes metric learning,further enlarges distances between different classes,ensures the validity of hash codes,and improves retrieval results.Finally,in this thesis,block encoding strategy is used to replaces the fully connection layer to reduce dimensions,and the slice layer is employed in deep hash methods of single labels and triple labels,respectively,which effectively improves the differences of features in different classes,makes the hash code more accurate,and improves the retrieval effect of remote sensing image.