Research On Retrieval Technology Based On Deep Learning

The research of multimedia retrieval technology is how to find effective data quickly and accurately.This technology has penetrated into many fields and provided great convenience for people's life and production.Therefore,how to retrieve information quickly and accurately has become an important issue in the research of multimedia retrieval technology.Retrieval technology has been effectively applied to public security,e-commerce,medical diagnosis,copyright protection and other aspects.Multimedia retrieval technology has become a hot topic in the field of computer vision and pattern recognition due to its important application value.In recent years,deep learning is widely used for large-scale multimedia data retrieval because it can obtain information better and improve retrieval efficiency.Although deep learning based multimedia retrieval technology has achieved good development,there are still the following problems: 1)Deep features information in image retrieval algorithm is not rich enough.2)The deep features is not fully used to learn semantic similarity.3)The rankng information of hash codes is not considered.4)The relative semantic similarity and multi-scale context information of images and sounds are ignored.In view of the above problems,this paper studies retrieval technology based on deep learning from four aspects.The main research contents and contributions are as follows:(1)Siamese dilated inception hashing with intra-group correlation enhancement for image retrieval.Because deep features information is not rich enough to enhance the intra-group correlation of hash codes,and it will cause the problem of reducing the correlation of similar hash codes.In this paper,from the perspective of features,the inception dilated convolution structure and category information are proposed to learn the multi-scale context information of features.Therefore,the proposed method can make use of rich feature information to enhance the intra-group correlation of hash codes.(2)Deep category-level and regularized hashing with global semantic similarity learning.Due to the insufficient use of deep features to learn semantic similarity,the similarity between hash codes is reduced.In this paper,deep features similarity learning is used to improve the similarity of hash codes.Moreover,the similarity of data is integrated in the process of learning deep features expression.As a result,the learned hash codes are more similar and discriminative.(3)Discrete deep hashing with ranking optimization for image retrieval.Ranking information of hash codes is very important to hashing technology.However,there is a problem that ranking information of hash codes is not fully utilized.This paper proposes to integrate discretization of hash codes and ranking information of hash codes into an overall framework,which can obtain discrete hash codes with obvious discrimination information.(4)Deep quadruple-based hashing for cross-modal image-sound retrieval.Because of neglecting the relative semantic similarity and multi-scale context information between images and sounds,the retrieval performance is poor.The algorithm uses the relative semantic similarity and multi-scale context information of images and sounds to improve the semantic relevance of images and sounds,which can improve the performance of cross-modal retrieval.