Pixel Club: Models of Stochastic Textures and their Applications in Image Processing

Ido Zachevsky (EE, Technion)
Tuesday, 25.7.2017, 11:30
EE Meyer Building 815

Textures are what differentiates true, real-life images, from cartoon images. The latter emphasize mainly smooth content other than the edges and contours, whereas the former stresses the importance of the details within the contours. Textures are found in facial images, natural scenery, aerial, medical and other types of images, and affect the image perception and recognition. Images with smoothed-out textures appear artificial and cartoon-like.

This study is devoted to a subset of textures known as Natural Stochastic Textures (NST). Stochastic textures are best modelled as random processes, as their repetitive structure is characterized in the statistical sense, and not in the spatial domain. After identifying the important statistical properties of NST, we propose several algorithms for texture super-resolution, denoising, deconvolution and other image processing tasks, using the fractional Brownian motion (fBm) as a suitable prior model.

Some NST cannot be fully described via a Gaussian model such as the fBm. This is due to the fact that Gaussian models are defined via their first and second order statistics and cannot, therefore, represent edges. Some NST, contain, however, additional structural characteristics that can not be modelled by fBm models. The complementary structural characteristics such as edges and thin lines is properly represented by the Fourier phase. We propose a suitable prior model for the local phase present in textures by incorporating complex wavelet decomposition into the combined model.

The analysis of textures that contain both stochastic and structural elements calls for texture embedding in low-dimensional spaces. We take a close look at the challenges involved in such embeddings. We conclude this study by proposing a method for texture analysis and show that the properties we propose in the first part arise naturally by learning the low-dimensional texture space.

Back to the index of events