Shadow Codes for Representation of Binary Visual Patterns

概要

論文の詳細を見る
In this paper, a novel approach to the representation of binary visual patterns is proposed, and the applicability of the method to recognition of handwritten patterns by neural network and conventional classifiers is investigated. The proposed approach has been named shadow codes, because it is based on shadow projections of pixels of the thinned input image onto the bars of a frame surrounding the image. A number of variations of the method can be devised, and the case in which the region of attention consists of a rectangle with oriention given by the principal axes of inertia of the input image is considered in detail. A frame composed by 16 bars classified into three categories is superposed on the attention region containing the thinned input image, and each pixel projects a shadow on the nearest bar of each category. While the determination of the attention region is inherently a translation-invariant process, scaling invariance is achieved by normalizing and quantizing the shadow lengths, resulting in a low-dimensional shadow vector. For a task consisting of the recognition of handwritten numerical characters using both a neural net, namely, a self-organizing map fine-tuned by learning vector quantization, and a conventional classifier, high recognition rates were obtained, confirming the effectiveness of the proposed representation method. Also, comparison with other graphical feature extraction techniques yielding feature vectors of the same dimension indicates that, although compact, shadow codes succeed in preserving information that can be used for recognition.
一般社団法人情報処理学会の論文
1998-02-15

Shadow Codes for Representation of Binary Visual Patterns

スポンサーリンク

概要

著者

関連論文

スポンサーリンク