Content-based Texture And Scene Classification

The local binary pattern is an effective texture descriptor.However,traditional local binary pattern operator only uses uniform patterns and ignores non-uniform patterns.As the radius and the number of sampling points increase,non-uniform patterns account for more and more percent of the total patterns.Discarding non-uniform patterns leads to much information loss.Attack to this issue,this paper proposes an improved local binary pattern operator.The operator extracts information from both uniform patterns and non-uniform ones.The improved local binary pattern operator divides non-uniform patterns into several categories according to the value of their uniformity.As such,this operator discovers an important set of local primitives such as lines,T-junctions,cross intersections and so on,which are as crucial as those represented by uniform patterns.Thus,the improved local binary pattern feature has more discriminative ability than traditional local binary pattern feature.Furthermore,ignoring the directions of all patterns results in rotation invariant,improved local binary pattern operator.Experimental results on the Brodatz and Outex texture datasets show that the improved local binary pattern method always outperforms traditional local binary pattern method.Due to large intra-class change,the accuracy of only using local binary pattern for scene classification is usually not high.Attack to this issue,this paper proposes a new,pyramid probabilistic latent semantic analysis algorithm for scene classification.The algorithm discovers topics of an image by using probabilistic latent semantic analysis,and embeds spatial information of topics by spatial pyramid.The algorithm also uses different kinds of descriptors,such as improved local binary pattern,scale invariant feature transform,histogram of oriented gradient,self-similarity,to descript local regions,and extracts different kinds of features which include complement information of an image.For classification,this paper employs AdaBoost classifier and prod-max decision fusion rule to get the label of test image.Experimental results on OT,LP,LSP,Caltech-6 and Caltech-101 show that the proposed scene classification algorithm achieves satisfactory results.