Rotation Invariant Texture Classification Based On Gabor Filters And Square Criterion Decision Tree

Rotation invariance is one of the most important characteristics in texture classification. This paper proposes an algorithm for rotation invariance texture classification. The algorithm estimates dominant orientation of an image based on Gabor filters firstly and then constructs a multivariate decision tree with square criterion. It is an accurate solution to rotation invariance of a texture image by estimating the dominant orientation (or main direction). Once the main direction is found using Gabor filters, the texture image could be adjusted accordingly, and then a feature extraction procedure is carried out to generate the feature vector. At the classification stage, multivariate decision tree is used as classifier for its speed and simplicity. But to build a decision tree, one has to search all possible combination of features and all possible separative hyperplane for the best split at each node of the tree. In this paper, we use the square criterion to achieve the best separation. The algorithm includes:1. Gabor filters based method to estimate dominant orientation. Gabor filters are widely used multidirectional anaylsis tools. Because these filters can draw out a lot of directional information from images, it is possible that dominant orientation is estimated with them. We conduct the comparing experiments between our method and the one based on traditional Radon transform.2. Square criterion to evaluate separative hyperplane. This criterion is an alternative to the conventional entropy criterion, information gain and Gini index. Furthermore, this criterion makes the searching of a hyperplane as simple as finding a solution to a linear equation set, instead of the traditional way’s tentative search, which takes heavy computational costs.3. Most-effective-search to select features. This method is designed to pick out the features that can effectively differentiate one class from another over the training set. With this method, we can overcome the drawbacks of traditional exhaustive search that requires tentative search on all possible combination of features. Thus the computational costs can be reduced greatly.We carry out experiments on both Brodatz database and CUReT database to evaluate our methods. The results show that our methods achieve high effectiveness in terms of computational costs with very little decrease in classification accuracy.