Vision Prize 2003 winner Mr Markus Turtinen
Paper Characterization by Machine Vision

Paper is a very difficult material for machine vision based quality control systems. Discrimination of different paper qualities by the human eye is not easy either. Paper formation and other quality measures are typically evaluated using "look through" techniques. These procedures are subjective and depend on different people's interpretations, and thus are prone to human error. To achieve more reliable quality control, evaluator independent methods are needed.

Different papers are visually highly similar even though they do not belong to the same quality class. This causes several requirements to the quality control system. Additionally, due to the large production capacity of a paper machine, quality control system must handle huge amounts of data. Fast processing of measured data is important in achieving a reliable and user-friendly quality control system.

The state-of-the-art paper inspection is often based on fl-radiation, but machine vision has also been used for measuring paper formation and quality. These methods are usually based on off-line measurement and suffer from poor accuracy. The developed new method is based on non-supervised clustering and effective texture features extracted from paper images. It can be adapted to both off-line and on-line paper characterization.

Using a non-supervised learning and clustering based approach for classifying the texture of paper, we can achieve more objective paper characterization. The new and powerful texture features are able to discriminate paper’s texture very effectively. Compared to the old, previously used features in paper inspection, like Co-occurrence matrices, FFT power spectrum analysis and specific perimeter, especially the Local Binary Pattern (LBP) method provided an excellent, even 40 times better classification accuracy than the traditional features.

The developed system is very simple to train and take in use. Some texture features of paper images are calculated and images are clustered with a self-organizing map (SOM) according to these features. Clustering is made in a non-supervised fashion and the need for human involvement in training is minimal. The method offers a very fast classifier and also a self-intuitive visual user interface. With the user interface we can visually find the areas in the map representing specific paper characteristics and search for the dependencies between the clustering of texture and real quality properties of paper.

The results attained in this research project were very promising and some industrial projects have started to develop the method to a complete on-line characterization system. The aim is to develop a fast and accurate characterization method, which can be easily attached to the automation and control system of a paper machine.