Texture Classification through Multiscale Orientation Histogram Analysis
Miguel Alemán Flores and Luis Álvarez León
Departamento de Informática y Sistemas, Universidad de Las Palmas de Gran Canaria, Campus de Tafira, 35017 Las Palmas, SPAIN
This work presents a new approach to texture classification, in which orientation histograms and multiscale analysis have been combined to achieve a reliable method for the comparison of textured regions. The gradients in every point of a texture are estimated and their orientations and relative magnitudes are combined to build an orientation histogram. We have used Fourier analysis to measure the similarity between the histograms generated by different textures, considering the effects of a change in the size or orientation to make our method invariant under these phenomena. Since different textures may generate similar histograms, we have analysed the evolution of these histograms at different scales, adjusting the filters which are applied to them when the multiscale analysis is carried out.
1. Orientation-Based Texture Classification
By means of the modified Newton Filters, we estimate the gradient and build an orientation histogram. The Fourier analysis of the histograms allows comparing the textures.
2. Texture Classification through Multiscale Analysis
Through the decrease of the gradients across the scales, we estimate the relation between the resolutions of two textures. Then, we compare the textures at different scales adjusting the std. deviation of the Gaussians which are applied.
The extraction of orientation histograms to describe the distribution of the orientations across a textured region permits us to perform an initial clustering of the textures according to the quantitative and relative distribution of the orientations. The comparison of the Fourier coefficients and certain normalisation processes which have been included allows a satisfactory classification in many cases, including size and rotational invariance. A further study has been carried out, by comparing the evolution of the histograms at different scales. We have extracted the scale factor which must be used when comparing two textures to perform the comparison appropriately. The quite promising numerical results obtained in the tests which have been implemented confirm the usefulness of the multiple comparison of the images, since they endow us with a much more robust discrimination criterion.
Scale Space - Skye (Scotland) - June 2003