Texture description and recognition using combined feature sets

The nonsubsampled contourlet transform (NSCT) overcomes the aliasing phenomenon of contourlet transform, and has better direction selectivity and translation-invariant than the contourlet transform which is essential for texture analysis. So according to the directional property and coefficients energy feature in NSCT decomposition, we proposed a new algorithm which combined NSCT with LBP. First, aligned two texture image through estimate the texture direction by Radon transform. Then, NSCT features are extracted from NSCT coefficients of images and LBP features is extracted in low frequency to realize the characteristic of multi-scale. Finally two features combined into a new feature named NSCT-LBP according to the different weight and we defined a new similarity measure for recognition. The algorithm is tested on texture images from Brodatz album and KTH-TIPS data set. Experimental results demonstrate that the proposed method produces more accurate recognition results than other well-known methods and overcome the deficiency of original algorithms in rotation and scale transform, and more robust to texture recognition.