Now using a random forest (instead of the earlier single decision tree or Learning Vector Quantisation method) which is trained with all classes. This method creates 65 decision tree based on random subsets of the training data.

Results in brief:

When comparing classification results to the corresponding ground truth segmentation images resubstitution of the training set into the classifier has a greater than 90% rate of correctly classified pixels whilst the testing set is around 70% - most images are classified above this rate but some significant failure cases lower this averagesuccess rate. This is based on a 39-image training set and 28 test images. Increasing the training set does improve results but takes some time.

Obviously we are more interested in the results from images unseen by the classifier in the training stage but it is still important that training images are correctly classified.

Initial Images and with respective overlays and result maps.