Insights into a new overclustering technique using machine learning for a self-selecting bin-restricted colour sorting setup for light red meranti (Rubroshorea spp.)

Timber colour sorting is an important woodworking process in producing a homogeneously coloured and pleasant looking product. However, for multispecific timber such as light red meranti, LRM (Rubroshorea spp.), which spans a wide gamut of colours, there is an antagonistic compromise between having good separability of colour and the number of bins. This research attempts to solve this by intentionally overclustering the intensity gamut and then automating the selection of ideal colour sorting bins (CSB) for a given batch size to produce high-similarity coloured sorting. 178,327 unique LRM wood samples collected over 8 months of production were used. Machine learning clustering algorithms such as k-means and Otsu multithresholding were tested against percentile and equal spacing methods. Batch sizes of 250 (B250) and 1,000 (B1000) pieces were evaluated. Maximum likelihood estimation was tested against statistical methods to select the CSB, and ideal overcluster setups were determined using the average delta E (\(\Delta E^*_{00}\)) assessment. The ‘burn-in rates’ of 3–30 pieces were then evaluated. For the B250 four-bin setup, six overclusters (6C4) performed best, with a recommended ‘burn-in rate’ of 12 pieces. For B1000, 5C4 performed best with a ‘burn-in rate’ of 10 pieces. The 4C3 configuration and the ‘burn-in rate’ of 10 pieces were found to be the best for three-CSB for both B250 and B1000. This study shows the feasibility of using machine learning to automate the bin selection process when the overclustering technique is used to improve colour sorting in situations with a restricted number of bins.