ニューロ識別における軸対称マスクの提案とGAによる最適化および最適結果の統計的解析

抄録

We have developed a neural network (NN) recognition and its hardware for paper currency. We have proposed mask concept to extract characteristics of the currency. Here, we propose a unique mask which has a symmetrical masked area against axes which divide a long side and short one of the currency. We can obtain the same value from both upright image and inverse one of the currency using the axis-symmetrical mask. This means that these values are invariant to upright and inverse of the currency conveyance. Namely, we can recognize twice number of currency kinds compared with the conventional mask though the NN construction is the same. Furthermore, we optimized the axis-symmetrical mask by genetic algorithm. Finally, we show the effectiveness of the optimization by statistical analysis of mask processed values.