Q22Digital Image Processing

Image segmentation is the process of partitioning a digital image into multiple segments (sets of pixels) to simplify its representation into something more meaningful for analysis, typically to locate objects and boundaries. Segmentation algorithms generally are based on one of two basic properties of intensity values: discontinuity and similarity.

Edge-Based Segmentation

Edge-based segmentation detects abrupt changes in intensity (edges) using first-order gradient operators like Sobel and Prewitt, or the optimal multi-stage Canny detector (Gaussian smoothing, gradient computation, non-maximum suppression, and hysteresis thresholding). Because raw edge detection produces disconnected pixel fragments, edge linking techniques assemble them into meaningful boundaries; local analysis links pixels with similar gradient magnitude and direction within a small neighborhood, while global techniques such as the Hough transform map edge points into a parameter space ($\rho ,\theta$) where collinear points produce intersecting curves, allowing lines to be detected even across gaps caused by noise or occlusion.

Region-Based Segmentation

Region-based segmentation, on the other hand, partitions the image into regions that are internally similar according to a set of predefined criteria (e.g., color, intensity, texture). Thresholding is the simplest form: a global threshold $T$ classifies each pixel as foreground or background based on $f(x,y)>T$, working well for images with bimodal histograms; Otsu's method automatically selects $T$ by minimizing within-class variance. Region growing starts from seed pixels and iteratively appends neighboring pixels that satisfy a homogeneity predicate. Region splitting and merging recursively subdivides the image into quadrants (using a quadtree) when a region is non-homogeneous, then merges adjacent homogeneous regions, balancing the over-segmentation risk of pure growing against the blockiness of pure splitting.

Comparison

Edge-based methods are fast and good at capturing fine boundary detail but are sensitive to noise and often produce broken contours requiring linking. Region-based methods are more robust to noise and guarantee closed, connected regions, but can be sensitive to seed selection and threshold choice, and may under-segment images with gradual intensity transitions. In practice, the two are often combined, for example using edge information to guide or validate region boundaries.

Watershed Segmentation

A third major category, watershed segmentation, treats the image as a topographic surface where pixel intensity represents elevation. Water is conceptually poured in from local minima, and as separate water basins are about to merge, a dam (boundary line) is built along the ridge, producing a set of closed, one-pixel-wide contours that naturally partition the image without requiring seed points or a global threshold. Applied directly to a noisy gradient image, this tends to produce severe over-segmentation, so in practice it is combined with marker-based initialization, where markers derived from region-growing or morphological preprocessing designate the true basins in advance, giving watershed segmentation a hybrid character that draws on both edge information (the gradient surface) and region information (the markers), making it one of the most effective segmentation techniques for touching or overlapping objects such as cells in microscopy images. Choosing among edge-based, region-based, and watershed approaches therefore depends on the noise level, the presence of touching objects, and whether closed boundaries are required by the downstream recognition task.