Abstract
This paper deals with a new approach to a high-speed pattern recognition of fish species, the shapes of which are very similar but the colors and/or patterns of which are different. The color features of fish are extracted from several small regions of a picture and are characterized by the co-occurrence matrix. Further, the classification of fish species is realized by the method of multi-level slice classifier. The relationship between the suitable number of the textural feature for fish classification and confidence interval of its parameters are clarified by experiments. The high-speed recognition time of 77ms per a fish can be achieved when position error is assumed to be ±5% in case of locating the small regions to extract the color features.
