Technical Approach
Proven low-level algorithms of co-occurrence classification and relaxation were used to segment the images into regions and edges. Unary and Binary features were then derived from the regions and edges. The features were utilised asinput to a feed-forward neural network for region classification. Control of the system was maintained by the use of a blackboard. As the net is a supervised classifier, the system operates in two modes, training and classification respectively. Training the network requires the creation of training data (ground truth). A classification tool for operator assistance in defining training data is based on a tree or hierarchy of region classes. The system is composed of the following components,
- Image segmentation into regions and edges
- Region and edge feature extraction
- Feature storage and maintenance in a database
- Region classification via a neural network
Data
While I have the data and results from my thesis, because I'm going to re-implement the software on the Mac, I've chosen a larger series of data to use. The data is still FLIR (forward-looking infra-red), is still unclassified, and still has the signal-to-noise ratio that I was using originally, but there's more of it. With the increased processing power at my fingertips, I think having more data is worth the pain of recreating the results.
The below image (together with its histogram) are typical of the data I'll be using.
| ← Contents | Segmentation → |