Modified oriented Gaussian derivative filter based texture detection algorithm and parameter estimation

Texture in images is used as a cue for various computer vision tasks like segmentation, classification and object detection. In this paper, we explore a variation of texture detection technique using oriented Gaussian derivative filters with multiple scales and orientations. After obtaining the filter responses at each pixel location, K-means clustering is used to determine regions with different textures. The modification to the algorithm is that, we are using only magnitude of the filter responses at each pixel location for determining regions with different textures through K-means clustering. This modification results in drastic changes in texture detection quality compared to original algorithm. In this work, we also propose a novel strategy to determine the optimal range of filter scale that yields semantically meaningful textures. The first and second order statistics of the filter response is used to determine the optimal range of filter scales. The proposed method is illustrated on synthetic images containing only textured regions with different textures.