Texture smoothing and object segmentation using feature-adaptive weighted Gaussian filtering

Abstract

Gaussian filter kernels can be used to smooth out textures in order to obtain uniform regions for image segmentation. In so-called anisotropic diffusion techniques, the smoothing process is adapted according to the edge direction in order to preserve the edges. However, the segment borders obtained with that approach do not necessarily coincide with physical object contours, especially in the case of textured objects. A novel segmentation technique by weighted Gaussian filtering is introduced. The extraction of true object masks is performed by smoothing edges due to texture and preserving true object borders. In this process additional features like disparity or motion are taken into account. The method presented has been successfully applied in the context of object segmentation in natural scenes and object-based disparity estimation for stereoscopic applications.