Maximum Entropy based Local Multiple Patterns for Texture Classification

This paper proposes a maximum entropy based local multiple patterns (LMP) to describe image textures. In traditional LMP, the thresholds have to be set empirically and the feature dimension is high. We adopt a maximum entropy based thresholding scheme, which computes the thresholds by dividing the intensity difference histogram of an image equally; We also use the split-concatenate encoding to form shorter and more effective feature vectors. Experimental results on four test suits with an SVM classifier show that our proposed method achieves overall better performances than both LBP and LMP in illumination and rotation invariant grayscale texture classification.