Image Complete Annotation Based On Convolutional Neural Network And Formal Concept Analysis

With the explosive growth of image resources on the Internet,how to use semantic tags to quickly perform image retrieval in the face of hundreds of millions of image resources has become one of the urgent problems to be addressed.Most of the traditional image labeling work is done manually,which is subjective and random,and it is easy to cause the phenomenon of missing image labels and incorrect labeling.Therefore,how to effectively annotate and enrich the semantic content of images has become a significant research topic.The convolutional neural network can better acquire the visual information of images through autonomous learning of image features because the complex feature fusion process is abandoned.Formal concept analysis is an effective semantic hierarchical analysis method.To effectively improve the semantic label of images,this paper studies the semantic complete automatic labeling of images based on convolutional neural network and formal concept analysis.The main work includes the following three aspects:(1)An image complete annotation algorithm based on convolutional neural network and semantic analysis of concept lattice is proposed.Firstly,this method uses the Image Net dataset to pre-train the VGG19 convolutional neural network model,and then uses the Corel5 k dataset to fine-tune the pre-trained model and saves the model;Secondly,the classification result of the input image to be labeled is taken as the initial label set,and the deep convolution feature after removing the softmax layer is saved.Then,a set of adjacent images is constructed,and a candidate set of labels is obtained by using concept lattice to analyze the semantic labels with high similarity and great relevance at the upper and lower levels.Finally,the experimental results on the data set Corel5 k show that the method can effectively improve the recall rate of image labels and achieve complete labeling of images.(2)In view of the different contributions of different semantic tags to image annotation,an image complete annotation algorithm based on convolutional neural network and weighted concept lattice is provided.Firstly,the deep convolutional neural network is used to obtain the initial tag set of the image.Secondly,the neighbor image set is constructed,and the weight of each label in the neighbor image set is calculated according to the information entropy theory to obtain the importance degree of each label attribute.Then,according to the calculation rules of semantic similarity of concept lattice,the redundant tag semantic keywords are removed and the candidate semantic tag words are improved.The experimental results show that this method can improve the precision of labeling to some extent.(3)A prototype system of complete image annotation is realized.According to the above research results,an image complete labeling prototype system based on deep convolution network and formal concept analysis is designed and implemented on Matlab r2016 b platform.