A1032
Title: An adaptive modal EM with application to image segmentation
Authors: Giovanna Menardi - University of Padova (Italy) [presenting]
Abstract: Image segmentation is the automated process of identifying distinct regions within a digital image for purposes as content retrieval, object detection, or pattern recognition. Digital images are composed of a fixed number of pixels, discrete elements encoding quantized values representing color or intensity. Segmentation involves assigning a label to each pixel so that those sharing similar properties, such as color, intensity, or texture, are grouped together. This goal naturally aligns with cluster analysis, which has become a central tool in image segmentation. Among clustering techniques, nonparametric methods are particularly well-suited for segmentation tasks, as they can identify segments of arbitrary shape and do not require the number of segments to be specified in advance. One of the most widely used nonparametric approaches is the mean-shift algorithm, a gradient ascent procedure based on kernel density estimation, where segments are associated with the modes in the color intensity and possibly spatial distribution. Kernel density estimators can perform poorly when modal regions vary in shape and size. Motivated by this limitation, an alternative nonparametric method is developed for adaptive density estimation and simultaneous group identification, which relies on a two-level EM-style algorithm: One layer for parameter estimation and one for mode detection. The aim is to investigate its application to the problem of segmentation in greyscale images.
