Image Semantic Automatic Labeling Method Based On Attribute Reduction

With the rapid development of internet technology, the popularity of digital cameras and the mature technology of image-sharing social networking have made the image data rise very quickly. In the database, semantic image automatic annotation is the key method of searching the valuable information quickly and efficiently. Bag of words model is an important method of expressing image. Since it can effectively express the content of image, BOV model has become a bridge between the high-level semantic and low-level visual. However, synonyms and polysemy in the initial visual dictionary in bag of words model have negative impacts in efficacy of image classification. Rough Set theory is one of the mathematical tools of reducing the attributes for imprecise, incomplete and uncertainty data. In this thesis, the reduction method of polysemy in the visual dictionary and the scene image semantic annotation method are deeply studied. The main research results are as follows:(1)A method of generate visual words based on the rough set attribute reduction was proposed for solving the problem that visual word ambiguity affect accuracy and efficiency of the classification in the traditional BOV Bag-of-visual words model. First of all, the training image set and visual dictionary were generated by using the BOV model, and were abstracted as decision information table. Besides objects in the decision table was label as decision attribute separately according to the class, visual words in the dictionary was label as the condition attribute; Then incompatible object equivalence sets were established according to decision attribute in decision table of equivalent set and equivalent condition attribute set, and the necessary visual words collection were generated by using heuristic learning for each condition attributes of decision table, keeping to the vision of the change of the incompatible equivalent set number words. Secondly, ambiguous visual words in the visual words package were eliminated effectively by measuring the importance degree of decision making unnecessary visual words in the table to save high importance value of visual words according to the necessary visual words collection and relative knowledge granularity, eliminating the low visual word of attribute importance, forming of reduction of visual word set. In the end, experimental results validate the effectiveness and feasibility of the method.(2)A method of image scene classification based on the rough set attribute reduction is presented. The visual dictionaries, generated by training images, are firstly constructed from the decision table by using the rough set theory. Secondly, the discernibility matrix is built according to different compressible sets composed of decision-making attributes, and the identification visual words are computed and retained. Then, the important degree of the visual words which is difficult to identify, is judged by the heuristic strategy to preserve the high importance of visual words and remove the low importance ones. And accordingly the visual dictionaries which can effectively describe the scene are generated,then obtain the rules of classification. Finally, experimental results show that our method can improve the performance of scene classification.(3)On the basis of the above research, an image semantic hierarchical annotation prototype system based on rough set is designed by using MATLAB as development tools.