CV Techniques

Fourier Transform – What and When

What:

-            Converts image from spatial domain to frequency domain

 

When:

-            Removing periodic noise

-            Analysing texture

-            Fast convolution (via FFT)

Fourier Transform – Why and Limitations

Why:

-            Efficient for large kernels

-            Reveals global frequency structure

 

Limitation:

-            Loses spatial localisation

Background Subtraction – What and When

What:

-            Separate moving subject from static background

 

When:

-            Gait silhouette extraction

-            Removing background clutter before segmentation

Background Subtraction – Why and Limitations

Why:

-            Simple and computationally efficient

-            Good for detecting motion

-            Clean silhouettes

 

Limitations:

-            Illumination changes

-            Dynamic backgrounds

Temporal Averaging

-            Good for stable scenes,

-            fails with illumination change

Spatiotemporal Averaging

-            more robust,

-            smooth noise across space and time

Temporal Median

-            best for removing intermittent motion,

-            robust to outliers

Mixture of Gaussians – What and When

What:

-            Models each pixel as multiple Gaussians

 

When:

-            Outdoor gait videos

-            Waving trees

-            Shadows

Mixture of Gaussians – Why and Limitations

Why:

-            more robust than simple averaging

-            handles dynamic backgrounds

-            silhouette extraction in gait

 

Limitations:

-            computationally heavier

-            Sudden illumination changes

-            If person stands still, absorbs into background

Template Convolution – What and When

What:

-            Apply filters

 

When:

-            Edge detection

-            Template matching

Template Convolution – Why and Limitations

Why:

-            Simple

-            Fast

-            Deterministic

-            Robust to small noise

 

Limitations:

-            Requires alignment

-            Computationally expensive

-            Sensitive to illumination changes

-            Not invariant to deformation

Statistical Filters – Mean

-            Smooths noise

-            Blurs edges

Statistical Filters – Median

-            Removes salt-and-pepper noise

-            Preserves edges

Statistical Filters – Gaussian

-            Smooths while preserving structure

-            Good pre-processing for edges

Gabor – What and when

What:

-            Capture local frequency and orientation

 

When:

-            Fingerprint ridge orientation

-            Iris texture encoding

-            Face recognition

-            Image coding

-            Image restoration

Gabor – Why and Limitations

Why:

-            Excellent for texture-based biometrics

-            Good for noise and illumination changes

-            multi-scale and multi-orientation

 

Limitations

-            computationally expensive

-            Sensitive to misalignment and occlusion

-            High-dimensional feature vectors

Intensity and Spatial Processing – Histogram Equalisation

-            Redistributes intensities to make histogram uniform

-            Boosts global contrast

Intensity and Spatial Processing – histogram normalisation

-            Forces image to have specific mean and variance

-            Consistent brightness/contrast across samples

Intensity and Spatial Processing – scaling

Linear – linearly maps pixel values from original range to new range

 

Linear with clipping – same as linear but clamps extreme values to avoid amplifying noise

 

Absolute value scaling – takes absolute pixel intensities before scaling (when filters produce negative values that need to be converted to usable form)

Intensity and Spatial Processing – Histogram stretching

-            Expands narrow intensity range to fill full dynamic range

-            Enhances contrast

Edge detection – First and Second derivative

First = gradient

-            Detects edge magnitude and direction

 

Second = Laplacian

-            Detects zero-crossings à edge localisation

 

When

-            Extracting boundaries (iris, palm lines, silhouette)

-            Preprocessing for shape descriptors

Edge Detection Operators

Roberts – computes diagonal gradients using tiny 2x2 operator, for fast, simple, minimal cost

 

Prewitt – uses 3x3 masks to estimate gradients, basic, noise-tolerant in low-quality

 

Sobel – Like prewitt but stronger centre weighting, more robust gradient estimation, better noise suppression

Feature Extraction – Template Matching

-            Slides template over an image to find similarity peaks

-            Used to locate known patterns when alignment controlled

Feature Extraction – Hough transform

-            Detects parametric shapes via voting

-            Iris boundaries, pupils, palm lines even with noise or partial occlusion

Feature Extraction – YOLO

-            Real-time deep learning object detector

-            Fast, robust face/person detection in unconstrained environments

Feature Extraction – Active contours

-            Deformable curves that lock onto object boundaries

-            Used for precise segmentation when edges are smooth but noisy

Feature Extraction – Unet

-            Deep learning segmentation network

-            Used for high-quality pixel-level segmentation when background subtraction unreliable

Feature Extraction – Symmetry

-            Measures bilateral symmetry

-            Used because humans exhibit stable symmetry patterns that help recognition and covariate analysis

Basic Thresholding

-            Converts image to binary using fixed threshold

-            Used for simple segmentation when lighting is stable

Otsu Thresholding

-            Automatically finds threshold that maximises inter-class variance

-            Used for robust binarization when foreground/background intensities differ