Research on Visual Pattern Recognition with Hierarchical Neural Networks

为了开发下一代视觉模式识别系统的新设计原则，我们并行开展了各种研究，并对作者先前提出的核认知模型和选择注意模型进行了改进。新认知电子管我们通过引入新的网络结构和改进学习方法的发展，成功地提高了新认知电子学在模式识别中的健壮性。将改进后的新认知器用于手写字母数字字符的识别。(1)利用边缘信息实现了粗细不变的线提取。(2)弯曲提取细胞的引入极大地提高了学习和识别能力。(3)通过对新认知器C细胞层的对偶，实现了感受野内的非均匀模糊，提高了识别变形模式的能力。(4)通过在感受野内产生非均匀敏感度，可以自适应地调整感受野的形状。还可以提高区分相似模式的能力。(5)提出了一种新的纠错学习方法。与传统的有监督学习相比，在模式识别中同样具有较高的鲁棒性。(6)如果在学习和识别阶段使用不同的阈值进行特征提取，则采用赢家通吃的无监督学习可以获得较强的泛化能力。一个草书英文单词识别系统已经开发完成。该系统由选择性注意模型组成，在该模型中加入了搜索控制器和自动注意切换机制。选择性注意模型的原理已被证明对手写汉字识别和人脸识别也很有用。

Aiming to develop new design principles for visual pattern recognition systems of the next generation, we have performed various researches in parallel and have improved the necognitron and the selective attention model, which were proposed previously by the author.1. Neocognitron We have succeeded in increasing robustness of the neocognitron in pattern recognition by the introduction of new network architectures and the development of improved learning methods. Performance of the improved neocognitrons has been tested for handwritten alphanumeric character recognition.(1) Thickness invariant line extraction has been realized by a new line-extraction method that uses edge information.(2) The introduction of bend-extracting cells has greatly improved the learning and recognizing ability.(3) Non-uniform blurring within a receptive field has been realized by dualization of C-cell layrs of the neocognitron, and the ability to recognize deformed patterns has been increase.(4) The shapes of receptive fields can be adaptively adjusted by creating non-uniform sensitivity within a receptive field. The ability to discriminate similar patterns can also be increased.(5) A new method of learning by error correction has been developed. The same high robustness in pattern recognition can be obtained with much smaller effort than the conventional supervised learning.(6) Large ability of generalization can be obtained by unsupervised learning with winner-take-all process, if different threshold values are used for feature-extraction in the learning and the recognition phases.2. Selective Attention A cursive English word recognition system has been developed. The system consists of the selective attention model, to which search controller and automatic attention switching mechanisms have been added. The principles of the selective attention model has been shown to be useful also for handwritten Kanji recognition and face recognition.