Concept Lattice Attribute Reduction Algorithm And The Scene Semantic Annotation

Concept lattice is a tool for the data analysis and knowledge discovery.Frequent weighted concept lattice is a kind of weighted concept lattice, whosenodes satisfy the threshold given by the users. BOV model is one of primarymethods for expressing images, and also one of methods for scene semanticannotation. In this paper, the evaluation criteria for attribute reductionof concept lattice and its fast reduction algorithm is constructed. Besides, a newgeneration method of visual words for scene classification based on frequentweighted concept lattice is presented. The main research work can besummarized as follows:(1) The evaluation criteria for attribute reduction of concept lattice and itsfast reduction algorithm are constructed. The evaluation criteria about attributereduction of concept lattice is proposed from the viewpoint of concept intent’sattribute relationship of the parent-child node, and the correctness of thereduction results are proved which lay a theoretical foundation for simplifyingthe steps of attribute reduction of concept lattice. On this basis, this paper alsogives a fast reduction algorithm for attribute of concept lattice, andthe algorithm’s time complexity is o(n). Finally, an example is analyzed toverify the correctness and effectiveness of the algorithm.(2) A new generation method of visual words based on frequent weightedconcept lattice is presented. Through analyzing and reducing visual words bymaking use of weighted concept lattice, the paper presents a novel method ofbuilding visual words. First the formal context about the BOV of training imageis generated, and the weight value of visual word is acquired throughinformation entropy. Second, the frequent weighted concept lattice of BOVmodel is constructed for each semantic category according to the intentthreshold given by the user. Then, select the reduced visual words whichcontribute to scene category to construct visual words dictionary according tothe extent threshold, and further improve the precision and efficiency of label. Final examples show that it is effective and feasible.(3) A scene classification by using the k-nearest neighbor method based onfrequent weighted concept lattice is presented. Final the experimental resultsshow that it is effective and feasible by using Lazebnik dataset which owns15kinds of natural scene images.