A comparison of seven thresholding techniques with the k-means clustering algorithm for measurement of bread-crumb features by digital image analysis Original Research Article

The suitability of seven thresholding methods (six algorithms: isodata, Otsu, minimum error, moment-preserving, Pun and fuzzy; and a manual method) to consistently segment bread crumb images was investigated in comparison with the previously reported k-means clustering technique. Thresholding performance was assessed by two criteria: uniformity and busyness of the binary images. Crumb features (cell density, mean cell area, cell uniformity and void fraction) were computed for each optimal threshold on 135 bread slice images. Slight variations in threshold led to substantial variations in crumb feature values, with cell uniformity and void fraction being more sensitive than the others. The manual method was inadequate for quantification of cell uniformity and void fraction. The fuzzy, Otsu, isodata and moment-preserving methods yielded good and consistent binary images. Although the fuzzy method showed relatively higher amount of busyness than the other methods, it was able to perform well on images with large void areas.