摘要 |
PROBLEM TO BE SOLVED: To realize a technique for searching a large amount image database for an object to be searched for at high speed, and, in an attempt to overcome problems in prior art, to achieve a practically required search time by using feature value+machine learning named GIFT by an inventor.SOLUTION: A key point feature (value) that is a local image feature invariable by rotation or expansion/contraction is calculated with regard to some image by making it executed by a computer. (a) With regard to an image comprising a plurality of pixels, a key point is detected on the basis of the gradation pattern of the image. (b) With regard to the detected key point, a pair of pixels is selected at random on the basis of candidate coordinates of a plurality of pixels distributed in the periphery of the detected key point. (c) With regard to the selected pair of pixels, a difference in luminance between the pixels is calculated. (d) A feature vector in which the calculated difference in luminance constitutes an element is calculated. A K-D tree is generated on the basis of the calculated feature vector, and a search is made by the generated K-D tree.SELECTED DRAWING: Figure 3 |