Counterpropagation neural networks for trademark recognition

This work considers the possibility of using the connected component algorithm for segmentation to extract the features inherent in the studied objects and counterpropagation neural networks (CPN) for the learning capability for object recognition. Neural networks do not need any mathematical model to determine the system output depending upon the given inputs. Instead they behave as model free estimators and their output is closest to the already "learned" patterns. Neural networks have conventionally been used for a variety of automatic target detection, character recognition, control etc., but in the case of multiple integrated object matching, such as trademarks, solutions have yet to be found due to the complex mixture of graphics and texts comprised in the logo. CPN operates on the principle of closeness in the n-dimensional Euclidian space. Very encouraging results are observed.