In this thesis, we examined artificial neuronal networks for their possible application in color classification of yarn pictures. With the help of the MATLAB package and its neural toolbox we developed an application which can be used for demonstration purposes and for future investigation of classification problems. Our tool allows a flexible configuration of multilayer perceptrons. After a training phase with preclassified yarns, the network can be applied to unknown yarns. A graphical user interface with broad capabilities allows a flexible configuration of networks. Various feature sets, pixel filters and parameters for training and testing procedures can be chosen.

In order to increase the accuracy of classification, we added several enhancements to the neuronal network. One of the improvements is the use of a two-step procedure where different feature sets are used in parallel networks. By combining the two results, the classification specificity significantly increases. A similar effect was demonstrated using a non-neuronal strategy which narrows down the classification space by the inclusion of negative learning patterns into the training process. In our tests up to twenty different yarns were successfully classified with great accuracy using simple color information. The accurate identification of untrained yarns proved to be far more difficult. Even the use of additional features such as the covariance of pixel clusters could not significantly improve the results.

Neuronal networks can be used as potentially robust classifiers which are easy to implement from the software point of view. Despite their versatiliy and flexibility, the success of neuronal methods may be limited by the quality of the raw data from which the classification features are extracted. Our results indicate that a neuronal network which uses the given picture material does not achieve the high accuracy conventional industrial classifiers are capable of.