Segmentation of Ill-Defined Objects by Convoluting Context Window of Each Pixel with a Non-Parametric Function

Abstract

Taking cue from the fact of Context-dependence in Human Cognition process, in this work an image segmentation method is introduced where each pixel are classified into its object-class depending on properties of its neighboring pixels within a context-window-frame surrounding it. In brief methodological steps, the convolution of array obtained as intensities of pixels of a context window is done with weights obtained through a specific architecture of Artificial Neural Network after training. The result of convolution is utilized to define class of an object. The training set of pixels is selected judiciously considering exhaustive variety of context-types which includes pixels inside, outside, boundary of objects. This work also gives a novel approach for quantitative assessment of segmentation-efficiency for a segmentation process. Also the use of context-window appears to improve the segmentation process because of equivalence of this approach with those which use a combination of local texture and color for segmentation.