Special Associative Preprocessing Structures for Qualitative Feature Extraction

摘要

Existing pattern recognition and classification algorithms in computer vision require vast amounts of computations on input data. As a result, memory access time is a critical parameter in system performance. Tremendous parallelism in structure and algorithm is required for the system to operate in real-time. A preprocessing structure for qualitative feature extraction which meets these system requirements is presented. In general, the structural architecture consists of a cellular array of pixel-processors each containing an inherently parallel associative memory element. As such, memory access time is minimal and parallelism is maximized. By varying this basic structure with regard to interconnection and additional logic, specific structures result which are capable of extracting measures of specific qualitative features. In this thesis two specific structures are described which extract, respectively, the qualitative features of texture regularity and line trend. Applications of these structures are presented. Low-level simulation and performance estimates indicate these applications are viable and amenable to real-time operation. Suggestions for the development of structures which extract other features or multiple features are described.