Image Classification and Segmentation

Classification or segmentation of textural and spectral features with respect to the shape of a small element, density and direction of regularity is of importance in image analysis applications. The PYXIS C++ Code includes spectral and texture analysis techniques using ISEA3H transformation procedures such as:  
  • Noise Removal, 
  • Clustering, 
  • Edge Detection, 
  • Watershed, 
  • Image Histogram.
Classification based on the distribution gray level histogram of a one channel image.  Classification often characterized as local minimum-value cells which are also generally determined around high-frequency distribution.  

Clustering Pipeline Example 

MRSAP is applied to an original image to separate the spatial high-frequency noise. The lower spatial-frequency component image creates a smoother histogram with less local minimum-value points.  The histogram is then processed through MRSAP to the original image directly - a 1-D analysis. With each low-pass filter a class is created based on the new histogram.  Segmentation is completed after the classification generates the boundaries between the different classes.  Classification with multi-channel images is similar but involves more complex combinations of the different channels. Other algorithms for classification, including multi-level slice classifier, minimum distance classifier, and maximum likelihood classifier can follow similar pipelines.

Noise Removal Pipeline

Method 1 - A filter calculator process that operates by iterated over all cells and comparing neighbors cell values.  If m other cell values are within a distance of n from the target cell, the cell value is considered valid, otherwise, the point is replaced with the average value of all of its neighbors.

Method 2: MRSAP is applied to an image to filter out high frequency noise. In this example, two high frequency resolutions were extracted using MRSAP and the added remaining two low frequency layers were added together.