Research on Visual Pattern Recognition with Hierarchical Neural Networks

为了开发下一代视觉信息处理系统的新设计原则，我们集中研究了生物大脑视觉系统中的活动过程。我们采用建模的方法来解决大脑的机制，并提出了神经网络模型来解释与主动视觉相关的各种功能。我们也尝试设计视觉模式识别系统使用的建模研究的结果。我们同时进行了各种研究，并取得了以下结果。(1)神经网络模型有两个独立的通道处理形式和运动信息。该模型利用选择性注意功能解决了约束问题。(2)双眼细胞的神经网络模型。该模型包括远单元、近单元和微调单元。我们还提出了一个理论，可以估计从一只眼睛被遮挡的物体的深度，并已表明，从我们的理论得到的结果与心理实验相吻合。(3)具有非均匀感受野的眼动模型。(4)空间记忆的神经网络模型。该模型记忆外部世界的零碎地图，并能通过链式回忆过程回忆出大范围的地图。(5)训练neocognitron识别真实的世界中的手写字符。我们以前开发的neocognitron是一种模式识别系统，其结构来自哺乳动物的视觉系统。我们使用大规模的手写数字数据库（ETL-1）来训练neocognitron，并获得了高于98%的识别率。(6)相关矩阵存储器的理论分析。

Aiming to develop new design principles for visual information processing systems of the next generation, we have concentrated our research on the active processes in the visual system of the biological brain. We used modeling approach to solve the mechanism of the brain, and proposed neural network models explaining various functions related to active vision. We also tried to design visual pattern recongnition systems using the results of the modeling research. We have performed various researches in parallel and have obtained the following results.(1) Neural network model that has two separate channels processing form and motion information. The model can solve the binding problem by the function of selective attention.(2) Neural network model of binocular cells. The model includes far-cells, near-cells, and fine-tuned cells. We also proposed a theory that can estimate the depth of an object occluded from one eye, and have shown that the results obtained from our theory coincide with the psychological experiments.(3) Eye movement model with non-uniform receptive fields.(4) Neural network model of spatial memory. The model memorizes the fragmentary maps of external world, and can recall a map of a wide area by a chain process of recalling.(5) Training neocognitron to recognize handwritten characters in the real world. The neocognitron, which we have developed previously, is a pattern recognition system whose architecture has been suggested from the mammalian visual system. We trained the neocognitron using a large-scale data base of handwritten digits (ETL-1), and obtained a recognition rate higher than 98%.(6) Theoretical analysis of the correlation matrix memory.