Research On Deep Hashing Methods Based On Remote Sensing Image Retrieval

As satellite observation technology improves,the number of remote sensing images significantly and rapidly increases.Therefore,a growing number of studies are focusing on remote sensing image retrieval,moreover,it poses more challenges include longer retrieval time and higher space consumption.Therefore,hash techniques have begun to be used on large-scale remote sensing image retrieval.Hash technology uses a shorter binary hash code for retrieval,so it can greatly accelerate the retrieval speed and reduce the occupied storage space.However,the manual features obtained by the traditional hash methods would lose part of the retrieval accuracy,so researchers tried to obtain the features using the deep hashing method to optimize the retrieval effect and improve the retrieval accuracy.Existing deep hashing techniques can further improve the retrieval effects by sampling effectively and using metric learning methods.For the low accuracy of the current deep hashing method and cannot keep the distribution of hashing code in Hamming space,this paper proposes to improve the retrieval speed and accuracy.The main contributions of this article are summarized as follows:1.Proposes a metric learning loss function based on hard probability sampling method.Existing sampling methods include random sampling,hard sampling,and semi-hard sampling methods,which only consider the characteristics of the sample itself but not the sample distribution characteristics from the overall perspective,resulting in uneven sampling and poor network training effect.In view of the problem,this paper proposes the probability hard sampling method,conducts the probability calculation according to the overall distribution of the sample,and makes the hard sampling according to its probability distribution,which making the distribution of training samples more even and the training effect better.Meanwhile,in order to solve the problem that the margin between positive and negative samples of existing loss functions is too small,the margin based loss function and an efficient quantization loss are proposed.Experiments show that the training network based on hard probability sampling method converges faster and is more stable,while the margin based loss function more clearly distinguishes the samples and gets higher retrieval accuracy.2.Introduction to the designing of the deep hashing method based on proxy loss.Existing pair-based deep hashing methods are building larger tuples to better learn the overall features of the samples,it can optimize the network training effect,but can bring a large amount of time consumption in the process of constructing sample pairs.Proxy loss learns the overall feature distribution of the sample by proxy,which can reduce training time and improve retrieval accuracy.In this paper,we propose a proxy loss using the optimal value which can make a large backpropagation gradient of samples far from the optimal value and make a small backpropagation gradient of samples close to the optimal value,so as to train the network parameters more effectively.3.The proposed deep hashing method based on classification label.Deep metric learning method based on the visual content information of the image itself can control the hamming distance of the hash codes from the same class to be small while the hamming distance of the hash codes from different classes to be large.This method can learn the highly discriminating hash code of embedding space and complete efficient retrieval,but it cannot further meet the needs of image analysis and processing to classify the samples as retrieval the correct samples and the classification label information of images is not fully utilized.This paper use the label information to construct a new structure of hash code and object function is constructed using label information for network optimization,which can improve the retrieval precision at the same time won’t bring extra space and time consumption.Considering that the existing metric learning loss cannot fully learn the classification semantic label,this paper designed object function combining the metric learning loss and the classification cross entropy loss.Through sufficient training on remote sensing datasets,efficient and accurate retrieval results are obtained.