Research On Image Annotation Based On Deep Feature And Multiple Kernel Learning

With the rapid development of the Internet,image data is explosive growth,there are a large number of unlabeled image data,and the traditional manual labeling method costs a lot.How to label keywords to these massive image data has become an urgent problem to be solved.Automatic image annotation is the core content to solve this problem,which has been widely concerned by researchers.In order to solve the problem of image weak annotation,an image annotation method based on integration of deep feature and word relevance is proposed in this paper.Firstly,the symbiosis and semantic relationship between tagging words are used to solve the problem of image weak annotation,and then the depth feature of image is extracted by deep convolution neural network,which effectively avoid the problem of insufficient expression of low-order artificial features in traditional way.Finally,based on integration of deep feature and word relevance,the relationship between image feature space and semantic concept space is established to complete the task of image annotation.Considering the problem of unbalanced distribution of tagging words in image data set,and all kinds of image features have different representation abilities to various semantic concepts.In this paper,a multi-feature fusion image tagging method based on multiple kernel learning is proposed.Firstly,the oversampling method of synthesizing a few class samples is used to overcome the influence of a few classes on the annotation performance.Furthermore,some traditional visual features and deep feature of image are fused based on multiple kernel learning,and the complementary information of multiple features is combined to enhance the feature representation ability of image.Thus,the internal relationship between image and the labelling words is obtained.The two labeling methods proposed in this paper are tested and evaluated on three benchmark datasets respectively.The experimental results show that the annotation method based on deep feature combined with word relevance has better performance than other image annotation algorithms.And through the experimental comparison under the different label number of the given training image,it shows the effectiveness of this method to solve the problem of image weak labelling.At the same time,the experimental results show that the multi-feature fusion tagging method based on multiple kernel learning can express the semantic information of image more accurately and fully,and complete the automatic image tagging task effectively.