Research On Image Retrieval Algorithms Based On Deep Learning

Image retrieval technology studies how to retrieve the most similar image from the massive image data,which has been applied to the e-commerce,the public security,the urban planning and other aspects.This technology has become an important issue in the field of artificial intelligence research.At the same time,with the continuous development of aerospace technology and remote sensing technology,remote sensing image retrieval has become a hot research topic in recent years.Feature extraction and similarity retrieval are the most critical steps of remote sensing image retrieval technology.When massive high-resolution remote sensing data are being processed,the existing methods do not fully consider the multi-channel characteristics of multi-spectral remote sensing images,resulting in insufficient feature extraction.At the same time,they also fail to fully learn the multi-scale information of features and fail to capture more contextual information,resulting in the loss of feature information.In addition,the existing cross-modal retrieval methods do not learn the consistent correspondence between image and audio well.Therefore,this paper takes advantage of deep learning to learn high-level semantic information and the advantages of fast and low memory of hashing algorithm to conduct in-depth research and exploration on single-modal retrieval of remote sensing images and cross-modal retrieval between image and audio respectively.Specifically,the main work content of this paper includes the following three parts:1)We propose an unsupervised variational auto-encoder hashing algorithm based on multi-channel feature fusion(VAEH).Multi-Channel Feature Fusion(MCFF)is employed to acquire the feature information of image,which fully considers the multi-channel properties of the multi-spectral RS image.In order to enhance the discriminability in the local preservation mapping process,variational construction process and automatic encoder are added into the learning process of hashing function,in which the KL distance of the Variational Auto-Encoder(VAE)is used to constrain the hashing code.Experiments on two large public RS image data sets(i.e.SAT-4 and SAT-6)have shown that our VAEH method outperforms state of the arts.2)We propose an unsupervised variational auto-encoder hashing algorithm based on multiple-scale dilated convolution(MECH).A new context feature enhancement based on multi-scale dilated convolution module is introduced,which is composed of multiple dilated convolution with different dilation rate.This module can increase the receptive field of convolution and gain more semantic information from the feature,which can solve the problem that the multi-scale information of the image is not fully learned and more contextual information cannot be captured.The proposed algorithm is also evaluated on SAT-4 and SAT-6 data sets.The experimental results prove that the algorithm has better retrieval accuracy.3)We propose a cross-modal retrieval algorithm based on deep similarity learning(CRDSL),which extracts the feature information of image and audio respectively,and the model is trained based on the deeply similar network.The proposed algorithm can solve the problem of retrieval between image and audio,and it is tested on PASCAL statement dataset and Wikipedia dataset.The experimental results demonstrate the feasibility of the proposed algorithm.