Research On The Visual Word Reduction Method Based On Binary Discernibility Matrix

The images data is exploding on the internet, along with the era of big data. Faced with the more and more image data, it has been unable to meet the actual needs which is using conventional label image manually, so how to find a quick way to automatic label image has been an important research content. Currently, the scene classification is a hotspot in the automatic semantic annotation. The model of bag of visual words(BOV) is an important way to express the scene image content. However, the formation of BOV will cause several problems, such as redundant visual words, "polysemy" and "synonymous", and these visual words seriously affected the performance of the scene classification. The method of binary discernable matrix is an effective method of rough set attribute reduction. This paper studies the reduction of visual words and the scene classification by using binary discernibility matrix model with a combination of BOV. The main research work is as follows:(1)A reduction method of redundant visual words based on binary discernibility matrix is presented. First, different 0-1 information decision table and binary discernibility matrix are produced from all the training images by adjusting the normalized threshold value of ?; Then the number of 1's in binary discernibility matrix is regarded as the heuristic information for identifying nuclear visual words and important visual words,and these words are combined to construct the classification decision rules which are used for the description of image. Thereby the impact of redundant visual words on the image scene classification is reduced, and the accuracy of image scene classification is improved. In the end, the experimental results on OT8 image data verify that the method is effective.(2)A reduction method of polysemy visual words based on binary discernibility matrix is presented. Since the decision of the normalization threshold value of ? has large impact on the rule generation in the(1) method, the distinguish degree of decision rules will be decreased with the increasing of the capacity of visual words if excessive visual words are removed. So the 0-1 information decision table and binary discernibility matrix are produced for any two types of different training images; Then the union set is respectively computed on the reduction of visual words from one type of image and any other types of image according to the binary discernibility matrix reduction algorithm, and this set as the decision rule of this type of image, thereby reducing "polysemy" problem from any two types of images presented in BOV and form a stronger distinguish ability of decision rules. In the end, some experiments are made on the image data sets of OT8 and Fei Fei13, the results verified the effectiveness of our method.(3) An image scene classification prototype system is developed based on binary discernibility matrix. Based on the given research work(1), the image scene classification prototype system based on binary discernibility matrix is designed and implemented by using Matlab and Java.