Cheat-sheetFeature extraction, table of contents
Chapter 8.1: Introduction
- 8.1.1. perClass feature extraction
- 8.1.2. Feature extraction domains
- 8.1.2.1. Extract features in local image neighborhoods
- 8.1.2.2. Representing objects identified in an image
- 8.1.2.3. Representing a band in spectral data
- 8.1.2.4. Transform color space
8.1.1. perClass feature extraction ↩
perClass provides a simple-to-use feature extraction framework based on
sdextract command. The syntax makes it easy to understand where it works
on and what is being extracted. The first three parameters are always
present:
>> sdextract( data, domain, feature [, options] )
The first one is the data set, second the domain string (where we extract features from) and the third one the name of a feature we extract. After, other options specific to a feature extractor may be given.
8.1.2. Feature extraction domains ↩
8.1.2.1. Extract features in local image neighborhoods ↩
- Domain: 'region'
- Use-case: Characterize local texture, structure at different image scales
- One sample: Corresponds to one local image neighborhood
- Data set is: Still an image with one band per feature
- Read more in: "Feature extraction in local image regions"
8.1.2.2. Representing objects identified in an image ↩
- Domain: 'object'
- Use-case: Local-to-object fusion, object recognition.
- One sample: Corresponds to one object as defined by object labels
- Data set: Describes a set of objects typcally defined by connected component segmentation
- Read more in: "Object feature extraction and representation"
8.1.2.3. Representing a band in spectral data ↩
- Domain: 'band'
- Use-case: Classification of materials by point spectra or in hyperspectral images
- Meaning: One feature vector per each original spectrum; one or more features per spectral band (connected set of narrow wavelengths)
- Input sample: Is a spectrum with many spectral wavelengths; there is connectivity between neighboring wavelengths
- Output sample: Corresponds to the input spectrum
- Output data set: Reduces the number of features from many wavelengths to few band features
- Read more in: "Feature extraction for spectral data"
8.1.2.4. Transform color space ↩
- Domain: 'color'
- Meaning: Input is a pixel in a color image, output is a pixel in a different color-space
- Use-case: Robust color-based segmentation