UNIT 4-Feature Extraction
UNIT 4-Feature Extraction
FEATURE EXTRACTION
In image analytics, feature extraction is the process of capturing important characteristics of an object or
region in an image. These features help in recognition, classification, and matching tasks.
Feature descriptors must often be invariant to scale, rotation, translation, and noise.
Representation
Two primary ways to represent objects in an image:
Boundary Representation: Only the outer boundary is used. Good for shape analysis.
Region Representation: Entire object region is considered, including texture and structure.
Example:
A leaf outline would be a boundary representation; the full leaf with veins and texture is a region
representation.
Boundary Preprocessing
Before boundary-based descriptors (like Fourier Descriptors or Shape Numbers) can be applied, the
boundary of an object in an image must be cleaned, extracted, and standardized. This involves several
key preprocessing steps:
Noise Reduction
Real-world images often contain noise due to lighting, sensor errors, or environmental conditions.
Noise can cause false or broken edges during edge detection.
Apply smoothing filters like:
Gaussian Blur: Most common; applies a weighted average to nearby pixels.
This reduces sharp intensity changes caused by noise while preserving important structure.
Edge Detection
Edges are the most important indicators of object boundaries.
Edge detection highlights the locations where pixel intensity changes abruptly, typically indicating the
boundary of an object.
Normalization
Objects in different images may appear at different scales, positions, or orientations.
To make feature descriptors comparable, we normalize the extracted boundary.
Steps:
Translation to Origin
Scaling to Unit Size
Rotation to Standard Angle
Boundary Feature Descriptors
Boundary feature descriptors are numerical values or encoded patterns that capture the shape characteristics
of an object by analyzing only its outline (boundary) — not its interior (region).
They help in recognizing and comparing objects, especially when the interior details are irrelevant or
unavailable
Applications:
Good for recognizing digits, letters, or geometric shapes.
Fourier Descriptors
These are frequency domain representations of shape boundaries.
Treat the boundary as a complex signal:
s(n)=x(n)+j⋅y(n)
Apply Discrete Fourier Transform (DFT) to get frequency components.
Benefits:
Low-frequency components capture the general shape.
High-frequency components capture fine details and noise.
Statistical Moments
These describe the distribution of the boundary points mathematically.
Moments describe the shape's spatial distribution.
Examples:
Mean: Center of mass of the boundary.
Variance: Spread of boundary pixels.
Skewness and Kurtosis: Measure asymmetry and peakedness.
Topological Descriptors
Texture Descriptors
These quantify patterns or variations in intensity inside the region — especially useful in gray-scale or color
images.
Edge/Gradient-Based Features
Describe how intensity changes across the entire image.
Edge Histogram Descriptor (EHD)
Measures edge distribution in horizontal, vertical, and diagonal directions
Color Moments (in color images):
Mean, Standard Deviation, and Skewness of R, G, B channels.
Statistical Features
Describe global intensity distributions and relationships.
✅ a. Mean, Variance, Skewness
Basic statistical moments of pixel intensities.
✅ b. Entropy
Measures randomness or texture complexity.
High entropy = more complex image.
Keypoint localization
Orientation Assignment
Each keypoint is assigned a dominant orientation based on the gradient directions around it.
This makes SIFT rotation invariant.
Application
Widely used in image stitching, object recognition, and 3D reconstruction, medical imaging
Property Description
Scale-invariant Detects features at multiple zoom levels
Rotation-invariant Accounts for orientation change
Robust Works well with noise, blur, partial occlusion
Distinctive Each feature is highly unique for matching
Local Based on patches, not whole image