Map of Contents. Classical image analysis built from first principles: pixels and filtering, gradients and corners, matching patches, changing resolution, and fitting models robustly. Learned visual models (CNNs and beyond) live in Machine Learning MOC, and image generation lives in Generative AI MOC.
Fundamentals
- Image Representation and Indexing - Arrays, channels, data types, (x, y) vs. (row, column), slicing, and views vs. copies.
- Pixel-Level Operations - Grayscale, thresholding, colour masks, blending, and the uint8 overflow trap.
- Image Noise - Gaussian, salt-and-pepper, Poisson, and speckle noise, and which filter fits each.
Filtering and Gradients
- Cross-Correlation and Convolution - Flipped vs. unflipped kernels, border modes, and the FFT.
- Image Borders and Padding - Zero, replicate, and reflect padding, and the false edges zero padding creates.
- Gaussian Filtering and Convolution - Gaussian kernels, box filters, normalization, and zero-mean kernels.
- Median Filter - A non-linear filter for salt-and-pepper noise, its limits, and how independent masks make the real corruption exceed the label.
- Image Gradients and the Sobel Operator - Gradient magnitude and direction, which kernel finds which edge, and smoothing first.
Feature Detection
- Harris Corner Detector - Structure tensor, the response function, thresholds, and border artifacts.
- Non-Maximum Suppression - Keeping one detection per object, on score maps, corner maps, and bounding boxes.
Local Features
- Scale Space and Difference of Gaussians - Scale space, blob detection, DoG as an approximate normalized Laplacian, and octaves.
- SIFT - Scale- and rotation-invariant keypoints: DoG extrema, localization, orientation, and the 128-D descriptor.
- Feature Matching - Nearest-neighbour descriptor matching, the ratio test, and RANSAC verification.
Matching
- Template Matching - SAD, SSD, NCC and ZNCC, the intensity model, when to use which, a salt-and-pepper noise experiment, complexity, and limits.
- Multi-Scale Template Matching - Finding the same object at several sizes by shrinking the image, and detecting multiple instances.
- Integral Images - Constant-time rectangle sums for fast window statistics.
Geometric Transformations
- Image Warping - Homogeneous coordinates, translation, rotation, scaling, shear, affine maps, and forward vs. inverse warping.
- Homographies and Perspective Warping - Perspective transforms, the DLT estimate from 4 point pairs, rectification, and when a homography is exact.
- Cylindrical Warping - Projecting onto a cylinder for panoramas, the mapping formulas, and choosing the focal length.
Resolution and Scale
- Aliasing and Downsampling - Why a low-pass filter must come before subsampling.
- Image Interpolation - Linear, bilinear, and bicubic interpolation, and inverse mapping for resizing.
- Image Pyramids - Gaussian and Laplacian pyramids, REDUCE and EXPAND, and uses from search to blending.
Model Fitting
- Least-Squares Line Fitting - OLS vs. total least squares, the SVD solution, and sensitivity to outliers.
- RANSAC - Robust fitting by random sampling: threshold, number of trials, and multiple models.
Code and Projects
- Image Basics Snippets - Indexing, thresholding, colour masks, blending, and box filtering.
- Corner Detection Snippets - Sobel, Harris, 3 x 3 NMS, and corner overlays.
- Image Warping Snippets - Transform matrices, inverse warping, affine and homography estimation, and a cylindrical warp.
- SIFT and Matching Snippets - OpenCV SIFT, ratio-test matching, RANSAC verification, and from-scratch DoG and matching code.
- Template Matching Snippets - SAD and NCC matchers, Gaussian filtering, the multi-scale search, NMS, integral images, and noise generation.
- RANSAC Snippets - Synthetic lines, SVD line fit, and a RANSAC loop.
- Computer Vision Coursework CS558 - Coursework results using these techniques.
Related Domains
- Machine Learning MOC - Neural networks and learning methods.
- Generative AI MOC - Image generation, VAEs, and GANs.
- Code Snippets MOC - Reusable code across domains.
To Research / Inbox
(Drop new concepts or terms you want to research here as unlinked wiki-links)
Next topics: the Canny edge detector, Hough transform, homography estimation, Fourier transform basics.