Research on Image-text Fusion Retrieval Methods for Picture Books

With the increasing development of social technology, the image data on the Internet shows an explosive growth. In a large number of complex and different databases, how to make better use of image features to achieve simple and efficient real-time retrieval is a new challenge we face. In recent years, image retrieval has begun to develop towards multi-feature fusion and specific application scenarios, but the addition of most text features relies on the text annotation of image information, and uses tags or description semantics for retrieval. In this way, the image information not mentioned in the image annotation will not be retrieved, and the requirements for the data set are relatively high. This paper proposes an image retrieval feature enhancement method fused with text features. Based on the spatial distance score weight, the text features are combined with the extracted VLAD (Vector of Locally Aggregated Descriptors) features to obtain the feature vector describing the image content. By adding text features contained in the image itself, the retrieval effect of images containing text information can be improved, and the retrieval accuracy of occluded images can also be improved to a certain extent, so that image retrieval achieves a relatively satisfactory effect.