極形式表示に適用した特異値分解による回転画像の認識

概要

論文の詳細を見る
In the field of digital image recognition, the extraction of rotation-invariant features is left as an important problem. In this paper, the original images which are represented in the Cartesian plane are transformed into polar plane as pre-processing and recognition of rotated images is made by applying singular value decomposition to the transformed images for the shift-invariant method. Although complex-log mapping has been used for the pre-processing to the original images, at this stage the transformation to polar representation is employed. That is to say, singular value decomposition is executed for the transformed image, and eigenvectors and singular values of the transformed image are calculated. For the translation of the transformed image adjustment of the eigenvectors can obtain a list of the same positive singular values. Compared with the method with complex-log mapping as the pre-processing, this method is superior in the simplification of calculation and recognition ratio. Finally effectiveness in relation to recognition of rotated images is shown by conducting computer experiments. At this stage, the recognition ratio is computed by applying the error-back-propagation learning to a 3-layer hierarchical neural network with one hidden layer.
東海大学の論文