In processing microscopic images of fibrous particles, it is difficult to select adequate threshold value for binarization since the images usually have unimodal distribution of gray level. Inadequate threshold value brings undesirable spots on background or cut-off of fiber images. In this paper a new binarization technique was proposed, assuming that background had a normal gray level distribution and its capability was compared with two existing techniques, i.e., the discriminant analysis and the neural network method. The experimental result showed that the neural network method gave the threshold value closest to that made by human judgement when the network learned adequately. Our proposed technique was second to the best. However, the discriminant analysis did not work well. Although the network has to learn again when the characteristics of images change beyond some allowable range, the criterion of the allowable range is not clear. Therefore, our proposed technique is the most practical one at present.

View full abstract