Scene Classification Algorithm Based On Markov Random Field And Fuzzy Set Theory

With the rapid development and popularity of Internet technology,the image data in the network grow explosively.In face of mass daily growth of pictures,how to classify and retrieval these images rapidly and accurately,has become a research hot spot.In the classification algorithm,the BOW model approach can establish the mapping between the underlying visual features and the semantics of the high-level scene.But there are two major limitations to the model:on one hand,visual words lack clear meanings;and on the other hand,they are usually polysemic.In order to break these limitations,this thesis focuses on the scene classification algorithm based on optimized visual words as the main research content,and studies the algorithms of feature coding and context information extraction.And because visual words can provide more information for a particular image category,giving them higher weights can enhance BOW performance in image analysis.According to this idea,the thesis presents a classifier design algorithm based on SVM(Support Vector Machine)combined with visual word weights.The main works of this thesis are as follows:1)The algorithm of pyramid visual bag model based on fuzzy set theory is proposed.Based on the traditional visual bag-of-words model,this algorithm improves the membership matrix of FCM algorithm.According to the distance between the image block and the cluster center,the distribution mode is set,which can avoid the vagueness of the meaning of the key point caused by too far away from the word and ensure the accurate information provided by the word when the distance is close.Experiments on public data set prove that this algorithm has high accuracy and good classification performance.2)The visual word generation algorithm based on adaptive a priori MRF(Markov Random Field)is proposed.Based on Markov random field theory,the algorithm organically links the feature appearance similarity with the contextual semantic information together.Firstly,LDA is used to derive the semantic symbiosis information between visual words.Then,the weight of the boundary term in minimum-cut/maximum-flow algorithm is used to calculate the parameters that control the intensity of the neighborhood.This method can reduce the ambiguity of visual words to a certain extent and obtain more accurate visual words of image blocks.3)The algorithm for the design of SVM classifier based on visual word weight is proposed.The algorithm firstly obtains the classification results through the polynomial kernel function classifier and the radial kernel function classifier.Then a novel word weighting method is introduced Based on the difference prediction labels generated by the two classifiers.The weighted Euclidean distance function is used to calculate the similarity between the image and the training set.The experiment results show that this algorithm has high classification accuracy.In order to verify the classification performance of applying the proposed word package in the scene classification algorithm,experiments were conducted on four data sets.The data sets include natural scene images(FeiFeiLi-15 data sets),human behavior images(HB-6 data set),simple object classes(MSRC-14 data sets),and complex moving images(UIUC-8 data sets).The final accuracy of the proposed algorithm is 86%on the UIUC-8 data set,83%on the FeiFeiLi-15 data set,93%on the HB-6 data set and 91%on the MSRC-14 data set.Compared with the current highest classification performance,the proposed algorithm surpasses the classification results on both HB-6 and MSRC-14 datasets.Experimental results show that the proposed algorithm is an effective scene classification algorithm.