The Pattern Distance for Images

Pattern feature is the intrinsic feature of an image. Compared with intensity feature and shape feature, pattern feature is more robust and it can provide lower ratio of within-class variance to between-class variance. Therefore, we propose a novel image distance based on the pattern feature called image pattern distance (IMPD). First we convert the image to the 3D histogram with both pattern and spatial information. Then we apply the local earth mover´s distance (LEMD) with Hamming and Lp ground distance to measure the 3D histograms. Experiments on the FERET face database and Outex texture database show that IMPD improves the recognition rate significantly.