A fast approach to tuning an adaptive mask for texture segmentation

Local textural features, generally in terms of texture energy, are extracted by linear filtering of an image with a set of N-coefficient zero-sum and symmetric convolution masks. If the texture energy is defined as a sum of square rather than an absolute value of the convolution between the mask and the textured image, the order of the average over a window of size W and the convolution may be interchanged. As a result, the computation time may be reduced by about 2W/N for general adaptive mask approaches that require tens of thousands of iterations during the training.