Osman Gökhan Sezer

Using perceptual relation of regularity and anisotropy in the texture with independent component model for defect detection

This paper addresses the raw textile defect detection problem using independent components approach with insights from human vision system. Human vision system is known to have specialized receptive fields that respond to certain type of input signals. Orientation-selective bar cells and grating cells are examples of receptive fields in the primary visual cortex that are selective to periodic- and aperiodic-patterns, respectively. Regularity and anisotropy are two high-level features of texture perception, and we can say that disruption in regularity and/or orientation field of the texture pattern causes structural defects. In our research, we observed that independent components extracted from texture images give bar or grating cell like results depending on the structure of the texture. For those textures having lower regularity and dominant local anisotropy (orientation or directionality), independent components look similar to bar cells whereas textures with high regularity and lower anisotropy have independent components acting like grating cells. Thus, we will expect different bar or grating cell like independent components to respond to defective and defect-free regions. With this motivation, statistical analysis of the structure of the texture by means of independent components and then extraction of the disturbance in the structure can be a promising approach to understand perception of local disorder of texture in human vision system. In this paper, we will show how to detect regions of structural defects in raw textile data that have certain regularity and local orientation characteristics with the application of independent component analysis (ICA), and we will present results on real textile images with detailed discussions.

Face recognition with independent component-based super-resolution

Performance of current face recognition algorithms reduces significantly when they are applied to low-resolution face images. To handle this problem, super-resolution techniques can be applied either in the pixel domain or in the face subspace. Since face images are high dimensional data which are mostly redundant for the face recognition task, feature extraction methods that reduce the dimension of the data are becoming standard for face analysis. Hence, applying super-resolution in this feature domain, in other words in face subspace, rather than in pixel domain, brings many advantages in computation together with robustness against noise and motion estimation errors. Therefore, we propose new super-resolution algorithms using Bayesian estimation and projection onto convex sets methods in feature domain and present a comparative analysis of the proposed algorithms with those already in the literature.