IDENTIFICATION OF TRASH IN CANE USING MACHINE VISION

BILLET CANE, leaf, and tops form the overwhelming bulk of sugar mill feedstocks, but currently there is no accurate and reliable non-invasive method of directly estimating the amount of trash (combined leaf and tops) they contain, or distinguishing between the amounts of leaf and tops. This paper reports on a novel technique that can reliably and non-invasively distinguish between billet cane, leaf, and tops on the basis of high resolution colour imagery. The technique distinguishes between billet cane and both leaf and tops using detailed surface texture. The features distinguishing cane from leaf and tops are consistent across all cane varieties examined and do not require calibration for cane variety. Texture is not a good discriminator between leaf and tops. However, leaf and tops can be distinguished on the basis of hue since the differences in water content and degree of senescence which distinguish leaf and tops are associated with hue. Texture and hue based discriminators are then merged in a probabilistic framework to produce a combined classifier capable of achieving true positive identification rates between 80% and 90% and false positive identification rates below 10% for cane, leaf, and tops in unseen test images.