Image Classification Techniques

Image is continuous data. To obtain the variation within the continuous data similar to discrete data, image classification data is important. Image are classified in the classes as required. 

Terms used: Feature Extraction, Pattern Recognition *Spectral* (pixel based on its pattern of radiance measurements) and *Spatial* (pixels classified related to its neighboring pixels, complex) *Temporal* (change in pixels over time)

Types of Classification:

•     Unsupervised: Aggregation based on spectral clusters; knowledge of thematic classes none, 
• The machine does all the work; popular but limitations remain
• Interpretation based on  the spectral relations,
• Human analyst determines their usability (renames the clusters) and may adjust clustering parameters,
• Advantages: Missed training sites in supervised gets chance to be visible and assumptions not required

•      Supervised: Training sets and areas are provided; based on "looks alike"                                       (similar spectral characteristics), based on users inputs

                         stages:  ~Training~Classification~Output

• "Training Stage: important for the success for the classification"
• Prior knowledge is required for technician. Computer assigns each pixel, which it seems to belong to, on the basis of provided training sites.

     

Algorithms Used:

•     Minimum distance to Means Classification: Center Point theorized for each cluster (minimum distance from mean)

•     Gaussian Maximum Likelihood Classification: Based on probability of contours that the point belongs to that class.

•     Parallelpiped Classification: a box is drawn for including pixels 

Object Oriented Classification:

Scene based on homogenous image objects (referred to as patches or segments) based on multi resolution image segmentation process.

 

(Contents as rough study notes)