The Representation of Visual Feature Variables in Connectionist Networks (Computer and Information Science)

本研究为连接主义网络中变量表征的描述和分析提供了一个一般性的理论框架。定义了所有表示所处的空间，并根据该空间描述现有的表示方案。此外，空间被用来提出新的表达方式。表示的数学分析用于根据不同的属性比较不同的表示。这种分析导致了关于视觉处理的理论，并被用来确定哪种类型的信息最好由特定的表示编码，因此对于为给定的计算任务选择最佳编码是有用的。该研究表明，在视觉处理的早期阶段，视觉特征变量的表示选择不仅取决于所需变量的信息类型，还取决于用于从图像中提取特征的测量过程。一种调整曲线技术已经被开发出来，用于分析测量过程，从而确定作为测量过程函数的表示选择的约束。这个项目的重点是研究一种特殊的表示技术，称为rho空间表示。它与用于边缘分析的简单并行连接计算一起发展。rho空间与其他边缘表示相比有许多优点，包括：不需要局部边缘元素的阈值；一种与人类感知一致的连接的自然表征被创造出来；粗略和精细的方向信息表示都是可能的；而且，虚幻的轮廓，由人类视觉系统产生的类型出现。

This research provides a general theoretical framework for the description and analysis of representations for variables in connectionist networks. The space in which all representations lie is defined, and existing representation schemes are described in terms of this space. In addition, the space is used to suggest new representations. A mathematical analysis of the representations is used to compare different representations in terms of various properties. This analysis has lead to theories about visual processing, and has been used to determine which types of information are best encoded by a particular representation, and is thus useful for choosing an optimal encoding for a given computational task. This research shows that in the early stages of visual processing the choice of representation for a visual feature variable depends not only on the type of information about the variable that is desired, but also on the measurement process used to extract the feature from an image. A tuning-curve technique has been developed for analyzing the measurment process, and thus determining the constraints placed on the choice of representation that are a function of the measurement process. The focus of this project is to study a special representation techniques, call Rho-space representation. It has been developed along with simple parallel connectionist computations for edge analysis. Rho-space has many advantages over other edge representations including: no thresholding of local edge elements is required; a natural representation of connectivity is created which agrees with human perception; both coarse and fine representation of orientation information is possible; and, illusory contours, of the type produced by the human visual system occur.