Classification of rotated and scaled textures by local linear operators

In the paper, we are concerned with classification of rotated and scaled texture by local linear operators. Firstly, we rotate a local linear operator to generate a rotated set. Each member in the set is convolved with the texture to obtain an orientation- and scale-dependent feature vector. Secondly, we convert the vector to rotation independent by moving the maximum moving average of the vector elements to the first position and rotating the other elements with reference to the relative position to the maximum moving average. Thirdly, we eliminate the vector variance due to scale change by normalizing each element with the minimum moving average of the vector elements. The experimental result shows that the classification accuracy of using local linear operators is high, for example, a single Laws mask of 12 rotated operators may give 86.2% classification accuracy for ten classes problem. When two masks are used, the accuracy may be as high as 92.4%