Texture Classification Using Three Circular Filters
Shripad Kondra, Vincent Torre · 2008
A new method for texture classification is presented. The proposed method uses only 3 circular filters. Images are first filtered using these filters, then thresholded and averaged over two small neighborhoods. Universal textons are generated without learning from the training sets. 80 universal textons are used for each neighborhood. The feature space is reduced in one neighborhood by grouping into 4 bins. Each image is thus represented by a 2D histogram giving a 320 (80 times 4) dimensional feature vector (Model). Models are then trained with Support Vector Machines using chi2 kernel. The results are compared with state of art texture classification methods on 4 texture databases. The proposed method performs better than all previously proposed techniques on the KTH-TIPS database, despite using only 3 circular filters.