A blackboard-based system for learning to identify images from feature data
A blackboard-based system which learns recognition rules for objects from a set of training examples, and then identifies and locates these objects in test images, is presented. The system is designed to use data from a feature matcher developed at R.S.R.E. Malvern which finds the best matches for a set of feature patterns in an image. The feature patterns are selected to correspond to typical object parts which occur with relatively consistent spatial relationships and are sufficient to distinguish the objects to be identified from one another. The learning element of the system develops two separate sets of rules, one to identify possible object instances and the other to attach probabilities to them. The search for possible object instances is exhaustive; its scale is not great enough for pruning to be necessary. Separate probabilities are established empirically for all combinations of features which could represent object instances. As accurate probabilities cannot be obtained from a set of preselected training examples, they are updated by feedback from the recognition process. The incorporation of rule induction and feedback into the blackboard system is achieved by treating the induced rules as data to be held on a secondary blackboard. The single recognition knowledge source effectively contains empty rules which this data can be slotted into, allowing it to be used to recognise any number of objects - there is no need to develop a separate knowledge source for each object. Additional object-specific background information to aid identification can be added by the user in the form of background checks to be carried out on candidate objects. The system has been tested using synthetic data, and successfully identified combinations of geometric shapes (squares, triangles etc.). Limited tests on photographs of vehicles travelling along a main road were also performed successfully.