Research And Algorithm Design Of Automatic Image Annotation

Image annotation plays an important role in solving the problem of image information overload on the network.Users can retrieve classified images by image labels.In the field of image,the research on image annotation mainly improves the model effect by mining image content or tag information and other auxiliary information,but it tends to mine image content,and the research on tag and other auxiliary information is not in-depth enough.In this paper,the idea of recommendation,from the label information and other auxiliary information,to solve the image labeling problem.Starting from the classic algorithm of the recommendation system(matrix decomposition)and the emerging tool of the recommendation system(deep learning and knowledge mapping),the recommendation method is used to complete the image tag to assist image classification and image retrieval.In this paper,images are compared to users in the recommendation system and labels to commodities in the recommendation system,and the marked relationship between images and labels,as well as the content information of images and the relationship between labels are used for analysis and research.Firstly,the classical algorithm(matrix decomposition)of the recommendation system is described and analyzed,and the principle,advantages and disadvantages of the matrix decomposition method are introduced.Aiming at the problems of long training time and poor interpretability,the optimization rule of Bayes(BPR)is combined with matrix decomposition to make full use of the partial order relation between image and label to improve the training speed of the algorithm and enhance the interpretability of the results.Secondly,referring to the idea of combining recommendation system and knowledge map,the sub-model of obtaining image visual content information using CNN and the submodel of obtaining deeper semantic information using tag relation map are fused.Firstly,the visual content information data of the image is obtained by training the model on the pre-trained Res Net101.Secondly,according to the co-occurrence relationship between tags,a tag relation map is constructed,and the historical tag set marked with the image is used as the seed node to carry out the graph diffusion and obtain the in-depth semantic information.Finally,the semantic information learned is combined with the visual information of the image to comprehensively measure the tag list that more closely matches the image and complete the task of recommending multiple tags to the image.Finally,the BPR optimized matrix completion image tag recommendation algorithm and the CNN-image diffusion based image tag recommendation algorithm are verified on real data sets Image Net,NUS-WIDE and MSCOCO.Experimental results show that the algorithm proposed in this paper can solve the problem of image annotation and its effect is improved.