Viirya

"Contour and Texture Analysis for Image Segmentation"

By JITENDRA MALIK, SERGE BELONGIE, THOMAS LEUNG AND JIANBO SHI

--

This paper focuses on the image segmentation problem. It tries to partition grayscale images into disjoint regions of coherent brightness and texture. The novelty of the proposed algorithm falls considering both texture and contour for image segmentation.

They apply the approach of textons, the prototypes of linear filter outputs, to map each pixel to the texton nearest to its vector of filter response.

The adopt the approach to convolve the image with a bank of linear spatial filters. The oriented filterbank used in this work is based on rotated copies of a Gaussian derivative and its Hilbert transform.

The collection of response images by convolving with the filterbank is referred as the hypercolumn transform of the image and is taken as the tool for contour and texture analysis.

For contour analysis, they use the oriented energy approach to detect and localize composite edges in an image. At the maximum of the oriented energy, the value is kept and converted to a probability-like number P{con}.

By simple comparison of the texton distributions on either side of a pixel relative to its dominant orientation, this work estimates the texturedness of the region surrounding a pixel as P{texture}.

The product of (1-P{texture}) and P{con} is taken as the contour cue and is used to calculate the contour weight W{ic} between two pixels.

Given two histograms of texton of two pixels, texture weight W{tx} is defined in this work too.

Combining contour and texture weights, this work applies the normalized cut framework for partitioning pixels in an image.