Object recognition by neural networks
Wei Li · 1990
Two model-based object recognition techniques using neural networks are introduced in this thesis for invariant two dimensional object recognition. The first technique formulates the object recognition as a graph matching process in which the nodes represent feature points and the arcs represent relations between feature points. The process is then mapped to an energy minimization problem which is implemented by a Hopfield neural network. A new global technique is proposed to match all the objects in the input scene against all the object models in the model-database at the same time. A locating process including coordinate transformation, coordinate parameter clustering and averaging follows the matching process. This algorithm can identify and locate an object in any position and orientation. The test scenes could consist of an isolated object or several partially overlapping objects. The performance of the proposed technique is compared with that of a relaxation technique. In the second technique, object recognition is formulated as a clustering process in the feature space of objects. A neural network of cascaded Restricted Coulomb Energy (RCE) nets is constructed for object recognition. The new idea involves a number of RCE nets cascaded together to form a classifier where the overlapping decision regions in a previously learned network are solved by the next network. Similarities among objects which have complex decision boundaries in the feature space are resolved by this multi-nets approach. Referring to the ability of the system to recognize correctly a new pattern even when the number of learning exemplars is small, the generalization ability of a RCE net recognition system is increased by the proposed coarse-to-fine learning strategy. A new feature extraction technique is proposed here to map the geometrical shape information of an object into an ordered feature vector of fixed length which is the required form for input to this neural network. The proposed techniques are invariant to object changes, such as positional shift, rotation, scaling, illumination variance, variation of camera set up, perspective distortion, and noise distortion. Experimental results for the recognition of several objects are also presented. A correct recognition rate of 100% was achieved on both the training and the testing input patterns in this second approach.