Editorial

The FSA, founded in 1953, is a professional association for specialists within the field of automation technology. FSA operations cover all branches of industry. At the beginning of 2000 we had 1,779 ordinary members, eight honorary members and 45 patrons in the Society.

Vision Prize 2001 winner Matti Niskanen  

The appearance of sawn timber has huge natural variations. Normal wood grain and knots should be discriminated in all circumstances, but methods used in state-of-the-art inspection systems frequently fail in this task. The textbook methods for visual inspection aim at systems that are trained by showing samples. Selecting and labeling the samples is an error prone process that limits the accuracy that can be achieved The result is likely to be a frustrating everlasting training task rather than an accurate automatic visual inspection device.

The developed new approach relies on non-supervised training and does not require labeling of individual samples but uses a clustering method to discover whether the samples fall into a finite set of categories based on the similarity of their features. Thus, it is not affected by human errors in training. 

The first part of the figure is a SOM (self organizing map) trained with material flowing through an inspection station. The boards have been divided into blocks for which features are calculated. Based on feature vectors each of these regions is clustered to a SOM. Similar blocks clearly cluster close to each other.

After training, the SOM is ready for use. The dashed line in the figure of detection SOM is an approximated boundary between areas containing sound and defected wood regions. Next in the figure is shown a piece of board where regions clustered to the defected side of the border are highlighted. 

For each of the suspected defects a new set of features is calculated and they are used to cluster them to recognition SOM. The image of the recognition SOM in the figure is produced from detections obtained from a number of boards. Only one defect for each node is shown and similar defects are clustered close to each other. When we take a look at all the defects clustered to an individual node, we see that they are very similar even though they may have not been labeled into the same category by a human expert. 

The achieved detection and recognition results are good and considerably better than ones obtained earlier for the same test material with other approaches. SOM provides also an intuitive user interface by giving a concrete way of seeing the classification problem.