SORTER BUILDING METHOD, IMAGE CLASSIFICATION METHOD, SORTER BUILDING DEVICE, AND IMAGE CLASSIFICATION DEVICE

摘要

To obtain the number of divisions of discretization that is suitable for actual data.SOLUTION: The number of divisions into which the value of a feature quantity axis is discretized/divided is set, and frequency distribution data that indicates the frequency of set number in each division is generated for each of a plurality of feature quantity axes. With respect to misclassification teacher data when the whole dataset is classified, a corresponding division to a value of the feature quantity axis is specified for each feature quantity axis by a sorter that classifies images on the basis of a by-class occurrence ratio in each division per feature quantity axis, and the frequency distribution data is corrected so that the frequency of teaching class in the corresponding division is increased. The setting to increase the number of divisions, the generation of frequency distribution data conforming to the set number of divisions, and the generation of a softer based on the frequency distribution data are repeated until the result of classification of the whole dataset by the sorter satisfies a first standard. When the result of classification satisfies the first standard, the feature quantity axis is divided to the number of divisions fewer than the number of divisions for the softer that satisfies the first standard and with which the result of classification satisfies a second standard.SELECTED DRAWING: Figure 3