Rotation and scaling invariant texture classification based on Radon transform and multiscale analysis

In this paper, we propose a rotation and scaling invariant feature set based on Radon transform and multiscale analysis. Radon transform is used to project the image to 1-D space, and then the rows of the projection matrix are transformed by an adaptive 1-D wavelet transform, thus the feature matrix with scaling invariance is derived in the Radon-wavelet domain. Multiscale analysis is employed for the feature matrix, and the energy values at different scales are proven not only to be invariant under image scaling and rotation, but also to reflect the different energy distributions of the texture image at different scales. In the classification stage, Mahalanobis classifier is used to classify 25 classes of distinct natural textures. Using the testing image sets with different orientations and scaling, experimental results show that the average recognition rate for joint rotation and scaling invariance of our proposed classification method can be 92.2%.