1/25
Vocabulary terms and definitions covering core paradigms, signals, filtering, feature detection, transformations, and mathematical foundations for computer vision.
The Three R's of Computer Vision
A foundational triad representing the vision pipeline: Representation (mathematical modeling of visual data), Reconstruction (recovering 3D structure from 2D images), and Recognition (identifying objects, behaviors, or scenes).
Separable Signal
A 2D signal that can be factored mathematically into a product of two independent 1D signals, enabling faster sequential 1D processing.
System Linearity
A system property requiring that both superposition and scaling hold, satisfying the relation T[αx+βy]=αT[x]+βT[y].
Low vs. High Image Frequencies
The former describes smooth spatial transitions like gradients or skies, while the latter captures rapid spatial intensity changes such as edges, fine textures, and fine details.
Shift-Invariant System
A system whose output behavior depends solely on input signal properties rather than spatial location, ensuring identical responses regardless of coordinate shifts.
Linear Space-Invariant (LSI) System
A processing system that allows complete characterization via an impulse response, fast computational filtering, and direct analysis in the frequency domain by satisfying superposition, scaling, and spatial translation invariance.
Impulse Response
The specific output produced by an LSI system when stimulated by a Dirac delta function, fully characterizing the system.
Fourier Phase vs. Fourier Magnitude
The first component determines where frequencies align and preserves dominant spatial structure and shape, while the second determines the amount of each frequency present without preserving structural layout.
Naive Deconvolution
An inverse filtering method that reconstructs an unblurred image by dividing by the system frequency response H(ω), leading to severe noise explosion when H(ω)≈0.
Wiener Deconvolution
A restoration technique using the expected natural image spectrum to prevent noise amplification by weighting lower frequencies heavily and dampening higher frequencies.
First vs. Second Derivative Filters
The former excels at detecting broad gradient transitions with wider localization, whereas the latter pinpoints edge zero-crossings with high precision but suffers from extreme noise sensitivity.
Sobel, Prewitt, and Central Differencing Operators
Discrete convolution kernels used to compute directional intensity gradients along image axes for edge detection.
Laplacian of Gaussian (LoG)
A feature detection filter that applies Gaussian smoothing followed by a second-order differential operator to highlight uniform blob-like structures at local extrema.
Scale Normalization in LoG
The practice of adjusting filter weights across varying scale factors to keep absolute energy constant, preventing response values from decaying as scale increases.
Harris Corner Detector
A method that identifies interest points by analyzing local intensity variations, exhibiting rotational invariance and partial intensity invariance, but lacking scale invariance.
Feature Detector vs. Feature Descriptor
The former identifies repeatable, structurally unique points of interest (such as corners or blobs), while the latter computes unique numerical representations of the surrounding patch for matching.
Color Histogram Descriptor
A visual representation that aggregates the global distribution of color frequencies across an image but completely lacks spatial coordinate awareness.
Scale Invariant Feature Transform (SIFT)
A computer vision pipeline consisting of multi-scale extrema detection, keypoint localization, orientation assignment, and local patch descriptor generation.
Image Warping vs. Image Filtering
The former alters the coordinate domain through geometric operations like rotation or scaling, whereas the latter alters the range by modifying pixel values while preserving spatial positions.
Affine Transformation Matrix
A geometric mapping accounting for translation, rotation, scale, and shear that preserves parallel lines but does not preserve absolute lengths or angles.
Degrees of Freedom (DoF)
The count of independent parameters that can vary in a geometric model, such as 2 for translation (tx,ty), 1 for rotation (θ), and 8 for a planar homography.
Forward Warping vs. Inverse Warping
The former copies source pixels directly to destination coordinates, often leaving unmapped gaps or holes, while the latter queries destination coordinates backward into the source image to ensure complete pixel coverage.
Homography
A 3×3 projective matrix mapping points between two planes up to an arbitrary scale factor k, preserving collinearity while distorting angles, distances, and parallelism.
RANSAC (Random Sample Consensus)
An iterative parameter estimation method designed to fit mathematical models from data contaminated with heavy outliers, isolating correct matches during tasks like panorama stitching.
Minimal Singular Value in Homography SVD
The solution vector chosen when solving Ah=0 that minimizes the algebraic error subject to the non-trivial unit constraint.
Lens vs. Pinhole
The former focuses light rays to map each scene point to a sensor pixel with high light collection efficiency, avoiding the extreme dimness or diffraction blur inherent to the latter.