M. Gamassi, V. Piuri, F. Scotti, M. Roveri, Genetic techniques for pattern extraction in particle boards images, Proc. of the 2006 IEEE Int. Conf. on Computational Intelligence for Measurement Systems and Applications (CIMSA 2006), pp. 129-134, La Coruna, Spain, July 2006, ISSN 1-4244-0244-1.
Industrial informatics
Inspection of Particle Boards
The Inspection of Particle Boards project investigates machine-vision techniques for automatic quality assessment of melamine-laminated particle boards used in furniture and related applications. These products typically consist of an inner layer of compressed wood particles and external printed layers designed to reproduce wood textures or solid colours. Defects on the visible surface can generate significant costs when faulty boards reach downstream manufacturers, making reliable inspection an important part of production quality control.
Traditional inspection is generally performed by human operators, even though the task is repetitive, tiring, and difficult to maintain with constant accuracy on high-speed production lines. The project therefore develops non-invasive automatic visual-inspection systems capable of detecting both printing defects and mechanically induced surface damage. The considered anomalies include white spots, hickeys, contamination, colour variations, tears, scratches, bumps, blobs, and misalignments of the printed pattern.
A central challenge is that particle-board surfaces contain complex and highly variable wood-like textures. Directly distinguishing a small defect from natural streaks, knots, and colour variations is difficult. Moreover, factories may switch among many printing patterns during the same day, and the original digital pattern is not always available to the inspection system.
The proposed solution introduces an automatic pattern-extraction stage before defect detection. The system analyses the acquired board image to identify the repetitive printing matrix used to generate the surface. The recovered pattern is then tiled, including horizontally and vertically flipped variants, to reconstruct an ideal defect-free version of the inspected panel. Subtracting the real image from this reconstructed reference produces a difference image in which defects are more easily separated from the complex background texture.
The complete inspection pipeline includes image acquisition with a linear CCD scanner, noise reduction, pattern extraction, reference-image reconstruction, difference-image generation, feature extraction, and classification. Geometrical and graphical descriptors, including object size, area, shape ratio, and colour, are extracted from the detected regions and used to determine whether the board is defective.
Experimental evaluation was carried out using real wood patterns and simulated particle-board images containing representative printing defects. The system correctly recognised 98.4% of defect-free boards and 94.9% of defective boards, corresponding to an overall defect-detection capability of 96.6%. Most errors occurred when the printing pattern was estimated incorrectly or when very small defects were located within strongly streaked wood textures.
The original pattern was recovered exactly in 85.4% of the experiments. A one-pixel error in pattern length occurred in 12.3% of cases, while errors greater than one pixel occurred in only 2.3%. The experiments also showed that uneven illumination can interfere with row and column extraction, highlighting the importance of controlled lighting and image normalisation in industrial deployment.
A further research direction replaced conventional pattern estimation with a genetic optimisation approach. Each candidate solution represents the position and dimensions of a possible printing matrix. Its fitness is evaluated by reconstructing the full panel from the candidate pattern and measuring the pixel-wise difference between the reconstructed and original images. Because exhaustive exploration would require billions of evaluations even for relatively small images, genetic search provides a more practical optimisation strategy.
Three reconstruction and fitness strategies were studied. Simple tiling was computationally efficient but produced a highly selective fitness surface that hindered convergence. An adaptive row-based strategy improved robustness to positioning errors and recovered patterns in 78% of the tested panels. The fully adaptive method compensated for variations in both position and pattern size and found a candidate pattern in 98% of panels within 300 generations.
The genetic approach demonstrated strong pattern-extraction capability even when no prior information was available beyond the general pattern shape and when the pattern appeared in transformed configurations. Its main limitation was computational cost: the reported implementation was not yet suitable for real-time industrial operation and required further software optimisation.
Overall, the project introduced a pattern-aware inspection framework tailored to rapidly changing printed particle-board surfaces. By automatically learning the reference texture from the product itself, the system improves defect visibility and classification accuracy without requiring the original printing file. The resulting methods provide a foundation for continuous, objective, and non-contact quality monitoring in particle-board manufacturing.
Traditional inspection is generally performed by human operators, even though the task is repetitive, tiring, and difficult to maintain with constant accuracy on high-speed production lines. The project therefore develops non-invasive automatic visual-inspection systems capable of detecting both printing defects and mechanically induced surface damage. The considered anomalies include white spots, hickeys, contamination, colour variations, tears, scratches, bumps, blobs, and misalignments of the printed pattern.
A central challenge is that particle-board surfaces contain complex and highly variable wood-like textures. Directly distinguishing a small defect from natural streaks, knots, and colour variations is difficult. Moreover, factories may switch among many printing patterns during the same day, and the original digital pattern is not always available to the inspection system.
The proposed solution introduces an automatic pattern-extraction stage before defect detection. The system analyses the acquired board image to identify the repetitive printing matrix used to generate the surface. The recovered pattern is then tiled, including horizontally and vertically flipped variants, to reconstruct an ideal defect-free version of the inspected panel. Subtracting the real image from this reconstructed reference produces a difference image in which defects are more easily separated from the complex background texture.
The complete inspection pipeline includes image acquisition with a linear CCD scanner, noise reduction, pattern extraction, reference-image reconstruction, difference-image generation, feature extraction, and classification. Geometrical and graphical descriptors, including object size, area, shape ratio, and colour, are extracted from the detected regions and used to determine whether the board is defective.
Experimental evaluation was carried out using real wood patterns and simulated particle-board images containing representative printing defects. The system correctly recognised 98.4% of defect-free boards and 94.9% of defective boards, corresponding to an overall defect-detection capability of 96.6%. Most errors occurred when the printing pattern was estimated incorrectly or when very small defects were located within strongly streaked wood textures.
The original pattern was recovered exactly in 85.4% of the experiments. A one-pixel error in pattern length occurred in 12.3% of cases, while errors greater than one pixel occurred in only 2.3%. The experiments also showed that uneven illumination can interfere with row and column extraction, highlighting the importance of controlled lighting and image normalisation in industrial deployment.
A further research direction replaced conventional pattern estimation with a genetic optimisation approach. Each candidate solution represents the position and dimensions of a possible printing matrix. Its fitness is evaluated by reconstructing the full panel from the candidate pattern and measuring the pixel-wise difference between the reconstructed and original images. Because exhaustive exploration would require billions of evaluations even for relatively small images, genetic search provides a more practical optimisation strategy.
Three reconstruction and fitness strategies were studied. Simple tiling was computationally efficient but produced a highly selective fitness surface that hindered convergence. An adaptive row-based strategy improved robustness to positioning errors and recovered patterns in 78% of the tested panels. The fully adaptive method compensated for variations in both position and pattern size and found a candidate pattern in 98% of panels within 300 generations.
The genetic approach demonstrated strong pattern-extraction capability even when no prior information was available beyond the general pattern shape and when the pattern appeared in transformed configurations. Its main limitation was computational cost: the reported implementation was not yet suitable for real-time industrial operation and required further software optimisation.
Overall, the project introduced a pattern-aware inspection framework tailored to rapidly changing printed particle-board surfaces. By automatically learning the reference texture from the product itself, the system improves defect visibility and classification accuracy without requiring the original printing file. The resulting methods provide a foundation for continuous, objective, and non-contact quality monitoring in particle-board manufacturing.