Research On Image Retrieval Based On Perceptual Hashing

With the rapid development of the technology of digital media and computer network,the rapid increase of online images on the Internet,it has become increasingly difficult to quickly and accurately find the image of interest in the vast images.Therefore,how to automatically analyze and understand image information,and then quickly,accurately,and comprehensively retrieve and select the required content from the massive and complex images,and finally realize the application of analysis,induction,tracking,etc.,is a urgent problems to be solved in the big data environment.The research of perceptual hashing draws on the concepts and theories of multiple related fields such as hashing and multimedia authentication in the field of traditional cryptography,while ensuring the sensitivity of content and achieving the perceptual robustness of perceptual hashing,which provides the more efficient and effective solution.However,due to the large amount of images in a large-scale environment,the research on perceptual hashing has its own particularity and complexity,and the requirement of its performance is higher,which makes it difficult for existing image-aware hashing algorithms to deal with the requirement for robustness,differentiation,and compactness of image retrieval in large-scale environments,which become the main factors that hinder the further development of research and application of perceptual hashing technology.It can be seen that the research of the perceptual hash algorithm for image retrieval in a large-scale environment is a very challenging research problem.In view of the current problems of the perceptual hash algorithm for image retrieval,firstly,a hash optimization algorithm based on feature fusion is applied to image retrieval.The image retrieval process separates into online stages,and the offline stage.On the offline stage,the features of the images are extracted from two angles,and then the maximum correlation features are hashed and connected with respectively.In the online stage,the relevant maximum feature points obtained from the image are mapped into the binary code by the same method,and finally they are retrieved.In the experiment,the perceptual hash algorithm meets the requirement of similarity preservation,and has advantages over the single feature hash algorithm.Subsequently,a new deep linear discriminant analysis hash algorithm(DLDAH)is proposed.On the one hand,this method simultaneously achieves the purpose of increasing the difference between classes and reducing the difference within the class through linear discriminant analysis,thereby increasing the discrimination of visual vocabulary.On the other hand,the machine learning method used to learn the hash function on small-scale data can quickly hash the newly obtained image to generate a compact hash code,which greatly speeds up the speed of online retrieval.The experimental comparison with the existing methods proves the feasibility,effectiveness and superiority of the method.Through the research in this thesis,we have proposed a series of methods to effectively solve the problem of large-scale image perception hashing in image retrieval applications.But looking forward to the follow-up research,we still need to explore new methods and improve existed method,such as considering image feature learning,fast parallel algorithms and other aspects to improve the performance of the image retrieval system.