ランダムマスクを前処理機構に有する小規模ニューラルネットワークによる貨幣識別技術

抄録

Recently, many researches have been done to apply neural networks to pattern recognition and various simulation results prove the ability of neural networks. However, compactness, transaction speed, and cost of these products are important design factors when we apply neural networks to commercial products. In this paper, we propose a structure reduction method for neural networks concerning these factors. We adopt a slab-like architecture to the proposed method to extract characteristics of the inputs. By using the random masks, we can avoid the possibility of generating the same slab values even when the inputs are different. First, we show that the structure of a neural network can be systematically reduced by the proposed method for alphabets that are represented as binary data. Then we apply the proposed method to bill money and coin data that are represented by gray levels and show the effectiveness of the proposed method.