Supervised/Trained Segmentation

(Object-oriented Classification)

Technology description

Current state-of-the-art object-oriented classification techniques, such as the commercial software eCognition, require significant labour, through a trial-and-error selection of several interrelated parameters to find an appropriate segmentation. The segmentation process is time-consuming and operator dependent. This fact significantly limits the application potential of object-oriented classification in commercial use.

The present invention is a method of segmenting a digital image comprising the following steps:

  1. Initial segmentation: Performing a preliminary segmentation of the image into sub objects.
  2. Segmentation training: Defining a model object by selecting sub objects that define the model object.
  3. Automatically find optimal segmentation parameters: Use the fuzzy logic approach to analyze the optimal segment and determine the optimal segmentation parameters.
  4. Object re-segmentation: Apply the optimal segmentation parameters to segment the entire image.
  5. Object-oriented classification: Based on the optimal segments obtained from Step 4, determine final classification results.

This supervised segmentation technique provides increased speed and accuracy for object-oriented classification. It significantly increases the automation degree of object-based classification using eCognition.

Advantages

The novel features of the proposed method are:

  • Finds optimal segmentation parameters for eCognition effectively;
  • Provides an automatic solution for finding optimal segmentation parameters by training the software using a few sample objects;
  • Is significantly faster, more robust and easier to process than current technologies;
  • Reduces manual operation time and increases classification accuracy.

Stage of development


A prototype has been developed and tested successfully in the laboratory.

Intellectual property

  • Issued US Patent: US 8,233,712 – Methods of Segmenting a Digital Image

This technology is available for licensing.

Contact information

Office of Research Services, Industry-Government Services
University of New Brunswick
Phone: (506) 453-4674
partner@unb.ca