Digital Image Processing: Image Segmentation

Vocabulary practice flashcards covering fundamental and advanced concepts of digital image segmentation, edge detection, thresholding, region-based methods, Hough transform, watersheds, and clustering.

Image Segmentation

A process that partitions the entire spatial region $R$ occupied by an image into $n$ sub-regions ($R_1, R_2, \dots, R_n$) such that the union of all sub-regions equals $R$, each sub-region is connected and disjoint, each satisfies a uniformity predicate, and any two adjacent regions fail the uniformity predicate when united.

First-Order Derivative

The discrete difference approximation defined as $\frac{\partial f}{\partial x} = f'(x) = f(x+1) - f(x)$, which generally produces thicker edges in an image.

Second-Order Derivative

The discrete derivative defined as $\frac{\partial^2 f}{\partial x^2} = f(x+1) + f(x-1) - 2f(x)$, which exhibits a stronger response to fine details such as thin lines and isolated points.

Double-Line Effect

The characteristic response produced by the second derivative at ramp and step transitions in intensity, forming a pair of values of opposite sign on opposite sides of the edge.

Laplacian Operator

An isotropic second-order derivative operator defined as $\nabla^2 f(x,y) = \frac{\partial^2 f}{\partial x^2} + \frac{\partial^2 f}{\partial y^2} = f(x+1,y) + f(x-1,y) + f(x,y+1) + f(x,y-1) - 4f(x,y)$, used for detecting fine details and isolated points.

Gradient Vector

A vector pointing in the direction of maximum rate of change of image intensity at coordinates $(x,y)$, defined as $\nabla f = \text{grad}(f) = \begin{pmatrix} g_x \\ g_y \end{pmatrix} = \begin{pmatrix} \frac{\partial f}{\partial x} \\ \frac{\partial f}{\partial y} \end{pmatrix}$.

Gradient Magnitude

The magnitude of the gradient vector at $(x,y)$, given by $M(x,y) = \text{mag}(\nabla f) = \sqrt{g_x^2 + g_y^2}$, which reflects the edge strength and is often approximated in practice by $|g_x| + |g_y|$ or $\max(|g_x|, |g_y|)$.

Edge Direction

The spatial orientation of an edge at a given pixel, which is perpendicular to the gradient vector direction $\alpha(x,y)$ and calculated as $\phi = \alpha - 90^\circ$.

Zero Crossing

The location where the second-order derivative of an image changes sign, signaling the precise center of a ramp edge transition.

Marr-Hildreth Edge Detector

An edge detection technique that filters an input image with an $n \times n$ Gaussian lowpass filter, computes the Laplacian of the filtered image, and determines edge locations by detecting zero crossings.

Laplacian of Gaussian (LoG)

A filter formed by computing the second derivative of the Gaussian function $G(x,y) = e^{-\frac{x^2+y^2}{2\sigma^2}}$, yielding a characteristic 'Mexican hat' shape with a circular central lobe surrounded by a negative valley.

Canny Edge Detector

An edge detection algorithm designed to be optimal for step edges corrupted by white noise, adhering to three criteria: low error rate, well-localized edge points, and a single edge point response.

Nonmaxima Suppression

A ridge-thinning stage in edge detection that retains the gradient magnitude $M(x,y)$ if it is greater than or equal to both neighbors along the discretized edge normal direction, and suppresses it to zero otherwise.

Hysteresis Thresholding

A dual-thresholding approach using a high threshold $T_H$ and low threshold $T_L$ to eliminate false edge points, retaining pixels above $T_H$ as strong edges and including pixels between $T_L$ and $T_H$ only if connected to strong edges via 8-connectivity.

Global Thresholding

A segmentation technique that applies a single constant threshold $T$ across an entire image to partition pixels into object points ($f(x,y) > T$) and background points ($f(x,y) \le T$).

Otsu's Method

An optimum global thresholding algorithm that automatically determines the threshold $k^*$ that maximizes the between-class variance $\sigma_B^2(k)$.

Between-Class Variance

The variance metric maximized in Otsu's thresholding, defined as $\sigma_B^2 = P_1(m_1 - m_G)^2 + P_2(m_2 - m_G)^2 = P_1 P_2(m_1 - m_2)^2$, representing the separation between segmented pixel classes.

Separability Measure

A normalized measure used in Otsu's thresholding, defined as $\eta = \frac{\sigma_B^2}{\sigma_G^2}$, where $\sigma_G^2$ is the global image variance.

Variable Thresholding

An adaptive thresholding method where the threshold $T_{xy}$ changes locally across the image based on neighborhood properties, such as local standard deviation $\sigma_{xy}$ and local mean $m_{xy}$.

Moving Average Thresholding

A dynamic local thresholding approach for document processing where the threshold is updated line by line in a zigzag pattern based on a moving average of intensity: $m(k+1) = m(k) + \frac{1}{n}(z_{k+1} - z_k)$.

Region Growing

A region-based segmentation procedure that groups pixels or subregions into larger regions by appending neighboring pixels that meet specific similarity criteria.

Pixel Aggregation

The foundational approach to region growing that starts with a set of seed points and iteratively appends neighboring pixels sharing similar gray level, texture, color, or shape properties.

Region Splitting and Merging

A segmentation technique that repeatedly subdivides an image region into four disjoint quadrants whenever a predicate $Q$ is false, and merges adjacent subregions when they collectively satisfy $Q$.

Quadtree

A tree data structure utilized in region splitting and merging where the root node represents the entire image and each non-terminal node splits into four child quadrants.

Hough Transform

A global feature extraction method developed by Paul Hough in 1962 that maps image points to curves in a parameter space (e.g., $\rho = x\cos(\theta) + y\sin(\theta)$) to identify shapes through accumulator cell peaks.

Morphological Watersheds

A segmentation approach interpreting an image or gradient as a topographic surface where regional minima represent catchment basins and the boundary lines built to prevent merging water bodies define the segmentation lines.

Catchment Basin

A topographic region associated with a single local minimum such that a drop of water falling at any point within the basin would flow directly into that minimum.

Watershed Lines

The divide lines or crest boundaries separating adjacent catchment basins, representing the final segmentation lines in watershed segmentation.

Over-Segmentation

A common problem in watershed segmentation where noise and small local gradient irregularities generate an excessive number of tiny, unwanted regional minima and fragmented basins.

Watershed Markers

Predefined internal and external connected components used to specify the only allowed regional minima, preventing over-segmentation in watershed algorithms.

K-Means Clustering

An iterative clustering method that partitions data points into $K$ clusters by computing cluster centroids, assigning each point to the closest centroid, and recalculating centroids until the assignments stabilize.