Computer Vision Core Paradigms and Techniques

Vocabulary terms and definitions covering core paradigms, signals, filtering, feature detection, transformations, and mathematical foundations for computer vision.

The 3 R's of Computer Vision

The foundational triad representing the pipeline of vision systems:

• Representation: How visual data is modeled mathematically.

• Reconstruction: Recovering $3D$ geometry or structure from $2D$ images.

• Recognition: Identifying objects, behaviors, or scenes.

Separable Signal

A $2D$ signal that can be mathematically factored into a product of two independent $1D$ signals (e.g., the $2D$ unit step function). Processing these results in significantly reduced computational costs because operations can be performed sequentially in $1D$.

System Linearity

A system property requiring that both superposition and scaling principles hold mathematically. It must satisfy the relation $T[\alpha x + \beta y] = \alpha T[x] + \beta T[y]$.

Low Image Frequencies

Smooth spatial changes representing broad scene elements like gradients or a clear sky.

High Image Frequencies

Sharp, rapid spatial changes containing structural details, textures, or fine edges like grass or animal stripes.

Shift-Invariant System

A system whose output behavior depends only on the input signal characteristics, not the spatial position where it is applied. If the brightness or response changes dynamically based on coordinates, the system is shift-variant.

LSI System (Linear Space-Invariant)

A system that is simultaneously linear and space-invariant. Key features include allowing full characterization via an input/output relationship, enabling faster computational filtering, and allowing direct analysis within the frequency domain.

Impulse Response

The exact output sequence generated by an LSI system when a Dirac delta function is passed into its input. This response completely characterizes the LSI system's behavior.

Fourier Phase

Dictates where frequencies align and contains the dominant visual structure and shapes of an image.

Fourier Magnitude

Dictates how much of each frequency is present but carries less explicit spatial shape layout on its own

Naive Deconvolution

An inverse filtering method used to reconstruct a clean image from a blurred observation. In the presence of noise, dividing by the system frequency response $H(ω)$ causes severe noise amplification at frequencies where $H(ω) \text{≈} 0$.

Wiener Deconvolution

A robust deconvolution algorithm that utilizes the expected natural image spectrum to prevent noise explosion. It applies heavier mathematical weights to lower frequencies (where signal strength overpowers noise) and dampens higher frequencies (where noise dominates).

First Derivative Filters

Excel at detecting overall gradient shifts but offer wider localization.

Second Derivative Filters

Provide highly precise localization of edge boundaries (zero-crossings) but are highly sensitive to noise artifacts.

Central Differencing, Sobel, and Prewitt Operators

Digital edge-detection convolution matrices used to calculate directional intensity gradients across image columns (x) or rows (y).

Blob Detection & Laplacian of Gaussian (LoG)

An operator used to find blob-like uniform regions. It applies a Gaussian filter to smooth out noise followed by a Laplacian operator to compute local maximums.

Scale Normalization (in LoG)

The practice of adjusting filter weights across varying scale factors to keep the absolute energy constant. Without it, response values automatically diminish as the filter scale expands, preventing proper scale comparison.

Harris Corner Detector

A method used to locate corners by analyzing local intensity variations. It is fully invariant to rotation (eigenvalues do not shift), partially invariant to intensity changes, and not invariant to scale due to a fixed checking window size.

Feature Detector

Finds highly repeatable, structurally unique points of interest (corners, blobs, or edges).

Feature Descriptor

Extracts unique, quantifiable characteristics surrounding those points to reliably match them across different images.

Color Histogram Descriptor

An image profile summarizing global color distribution. Its primary limitation is that it possesses no spatial awareness (it does not capture structural layout or pixel positions).

SIFT (Scale Invariant Feature Transform)

A popular feature detection pipeline executed in four successive stages: - Multi-scale extrema detection: Identifying candidate keypoints across scales. - Keypoint localization: Refining keypoint positions. - Orientation assignment: Determining localized dominant gradient headings. - Keypoint descriptor generation: Encoding pixel patches into robust feature vectors.

Image Warping

Image Warping: Alters the domain (modifies pixel coordinates/geometries via scaling, rotation, or shearing).

Image Filtering

Alters the range (modifies actual pixel intensity values while maintaining spatial indices, like blurring).

Affine Transformation Matrix

A transformation matrix that can correct or induce scaling, rotation, translation, and shearing. It maps parallel lines to parallel lines but does not protect absolute distances or angles.

Degrees of Freedom (DoF)

The number of independent parameters that can vary within a geometric transformation model. For example, 2D translation has 2 DoF ($t_x, t_y$), 2D rotation has 1 DoF ($θ$), and a 2D homography matrix has 8 DoF.

Forward Warping

Iterates through source pixels and copies values to computed destination coordinates, frequently leaving structural 'gaps or holes.'

Inverse Warping

Iterates directly over destination coordinates and samples backward from source pixels, ensuring no pixel gaps are left unassigned.

Homography

A 3x3 projective transformation mapping points from one plane to another to alter visual perspective. It preserves straight lines but changes angles, parallel lines, and absolute distances. It is defined up to an arbitrary scale factor, meaning multiplying by a constant factor $k$ does not change output pixel coordinates.

Lens (vs. Pinhole) 

An optical component designed to angle incoming light rays to form a crisp 1-to-1 mapping from a physical scene point to a sensor pixel. It achieves high light collection efficiency while maintaining sharp focus, which avoids the severe blur or extreme dimness associated with simple pinholes.

 RANSAC (Random Sample Consensus) 

An iterative parameter estimation technique designed to fit models from noisy data containing significant outliers. In computer vision tasks like computing panoramas, it systematically isolates and removes incorrect keypoint matches so they don't corrupt the transformation estimation.

 Minimal Singular Value (in Homography SVD) 

When solving the stacked homogeneous linear system Ah=0, the vector paired with the minimum singular value is chosen to minimize the algebraic error ∥Ah∥22​ subject to the non-trivial constraint ∥h∥22​=1.