Research Of Artistic Image Captioning Technology Based On Transfer Learning

Adding up description text on art image simplifies the readers' understanding process while conduces to the artwork management.What matters is that the diversification of art cast significant barriers on its understanding,and the lack of the general art image caption dataset makes the task even harder.Till now,compared with the relatively mature Image Caption task,the research of art image captioning and labeling is in its initial stage.Meanwhile,artificial intelligence's progress performance in Perceptual Intelligence has shown its potential in Cognitive Intelligence,which leading people's chasing of artificial intelligence art.Therefore,this thesis takes art image as the research object to explore the ability of existing technology in understanding art image.This thesis proposes a transfer learning-based art Image Caption technology,which can automatically generate a readable content description for an art image.Firstly,we create a high-quality and high-precision art image description data set,Art Work.This Art Work dataset contains 1779 pairs of images and captions,and these data are annotated with text analysis tools afterward.A text content similarity assessment model is built to filter the generalized datasets' content based on these annotations.Secondly,combined with the development of image understanding,the framework of art image captioning is built,the main body of which is the image description generation model.Macroscopically,this thesis uses the structure of encoder and decoder to generate annotated text from the image,comprehensively considers the results of image feature acquisition and text vector embedding,and cast special attention on the mapping process between the image feature and the text content,adopts an efficient mechanism based on dual attention.The strategy for involving transfer learning in art image caption is analyzed,especially for its application in model training and data reconstruction.Through transfer learning,the knowledge of the relationship between image and text of the general image data set is fully used,and the differences between the content of different data sets are adapted.By casting practice on the MS COCO dataset and Sem Art dataset,the art Image Caption method we proposed in this thesis is evaluated.The practice has proved that the proposed model achieved a rise from 22 to 55 on the Sem Art dataset.All these practices have proved that the captioning technology proposed in this paper can improve the ability of art image understanding.However,limited by the lack of large-scale art image understanding data set,there is still a long development potential.