Research And Implementation Of Key Technologies For Large Scale Multi-modal Image Retrieval

With the popularity of the Internet and the progress of information technology,compared with the text,people are increasingly inclined to use images to express,transfer and gain information.As a result,the number of images on the Internet presents an explosive growth trend;the application of large-scale image data is facing great challenges.How to retrieve similar images from massive image data quickly and accurately is one of the hot topics in the field of image retrieval.Machine learning is one of the most important tools for dealing with large-scale data.Hashing learning has become a hot topic in recent years because it can reduce the storage space of data and its computational performance.Hashing technique maps high-dimensional data to low-dimensional Hamming space and obtains compact binary hash codes,which can reduce the storage space of data.At the same time,hash codes can keep the similarity between original spatial data.When searching,we can quickly get the similarity between the data by calculating the Hamming distance between hashing codes,thereby retrieval efficiency can be increased.In this paper,a Discrete Multi-view Hashing(DMVH)method based on Spectral Hashing for multimodal image retrieval is proposed.This method can improve the retrieval performance by using multi-modal feature-rich information.Firstly,the multimodal features(such as GIST,SIFT)are extracted and preprocessed to make the multimodal features consistent.Then,a new method to construct the similarity matrix is proposed,which not only preserves the local similarity of data,but also preserves the semantic similarity between the data.While preserving the similarity between data,the mapping matrix is used to map the data of the high-dimensional space to the low-dimensional space to obtain the hash code.As the hashing code is discrete,it is difficult to optimize the method directly.Based on this,two auxiliary variables are introduced,so that in the process of optimization,the discrete conditions are not relaxed to obtain higher quality hash codes.In this paper,the validation experiments compared with several advanced hashing methods are performed on three public datasets.Experimental results show that the proposed method outperforms the hashing method we compared with.Finally,based on the method proposed in this paper,a multi-modal image retrieval system based on discrete hashing is designed and implemented.The system is mainly to provide users with "image search image" function.After features are extracted,the user uses the learned hash function to map image to hash code.And then just calculate Hamming distance between the learned hash code and hash code in the database.The image corresponding to the smallest Hamming distance will be returned to the user as search result.