Cross-modal Retrieval Algorithm Based On Target Detection And Graph Convolution

With the advent of the data age and the information age.Information and data have increasingly become an important driving force for social and economic development and improvement of people's lives.Retrieval is particularly important as an efficient way to obtain information.As an important method for obtaining cross-modal information,cross-modal retrieval has huge social value,and naturally attracts more and more people's attention and research.With the development of deep learning and artificial intelligence,cross-modal retrieval has also made great progress.Among them,the deep learning method combining target detection and graph convolution has attracted more and more attention,and has become an important research direction in cross-modal retrieval.This article also pays attention to this direction,and conducts research in this direction from the following three aspects.1.Develop the learning ability of different layers in the multi-layer graph convolution to enhance the learning ability of the multi-layer graph convolution.The application of graph convolution in cross-modal retrieval is often continuous multiple layers,so that it can play a better effect than a single layer.However,in the past,only the output result of the last layer convolution was used,and the three-dimensional development of other layers of the multi-layer image convolution was not used.In this paper,we use the jump design mode to jump the middle layer through the subsequent picture convolution and directly enter the following process to achieve flexible control of the different layers of the multi-layer picture convolution,enabling different levels of learning and development of different layers.The results show that the method in this paper can improve the feature learning ability of graph convolution as a whole and improve the effect of cross-modal retrieval.2.Through multi-granularity text feature learning to improve the feature learning ability of text part.In the previous cross-modal retrieval,the feature learning of text is accomplished by using a simple cyclic neural network(GRU).Such a learning method of text features is too simple to fully learn the text information.Instead of GRU,we use multi-granularity text feature learning.Through multi-granularity text feature learning,the learning ability of the text feature learning part is enhanced.The text part can learn rich multi-granularity text information and fully learn the information in the text.The results show that multi-granularity text feature learning enhances the ability of text feature learning and improves the anti-interference ability of retrieval.When the amount of retrieved data is larger,the retrieval effect will decline relatively less.3.Improve the overall retrieval effect through mixed retrieval.In the past,cross-modal retrieval has always built a retrieval model in the algorithm and used the retrieval capability of the model to carry out the retrieval.Inspired by hybrid recommendation algorithms.We construct two retrieval models in an algorithm framework,make the two retrieval models work at the same time,integrate the two retrieval models together in a certain way,and use the overall retrieval capability of the two retrieval models for retrieval.Because each retrieval model has certain retrieval ability,when these models are superimposed together,the retrieval ability will be enhanced mutually,and the overall retrieval effect will be better.When the state consistency of the feature vectors of the two models is high in the vector space,the enhancement effect will be more obvious.As the two retrieval models work together,the anti-interference will be enhanced.When the data size of retrieval is larger,the retrieval effect will decline less.This paper focuses on the research on the cross-modal retrieval algorithm combining target detection and graph convolution from three aspects of multi-layer graph convolution middle layer using,multi-granular text feature learning,and hybrid retrieval.And experiments show that the improved methods and theoretical innovations in this paper have improved the retrieval effect to a certain extent.