TLS Grammars : Inference, Parsing, and Some Results

Abstract

AbstractThis paper describes a grammatical 2D-pattern recognition process. In order to build the in­ ference process on stable features of the object being studied, we set up a new type of grammars, so-called TLS-grammars. Attributes and related operators on chains, which are defined in this 2D-application as particular links between confirmed segments, are introduced inside the grammar productions and processed within the parsing step. The grammar rules are weighted with probabil­ ities; they are dynamically modified during the parsing step. Finally, experimental results from real patterns are discussed.Keywords: multivocal function - formal languages - regular and context-free grammars - Barley's algorithm - stochastic parsing. 1 Introdution and background Formal languages are a very attractive tool to perform real pattern recognition, since they work on the structural part of the information, giving programmers the possibility to know the why and the who about the decision of a recognition step (a good overview to syntactic pattern recognition is [Mfa85]).Nevertheless, they also have their own limits, for it may be difficult to determine which language is associated with which source objects; and furthermore why has a given language been inferred instead of another.Very often, syntactic methods need help from stochastic models to perform the recognition of real patterns - and in the same manner, attributes (which are in fact additionnal structural information) are set on the measures used in stochastic methods.This paper describes a new syntactic method for pattern recognition. It was primarily developed in order to process very similar shapes like motor-trucks or military vehicles. In this application, the problem was to find significant differences between such subjects. More precisely, the differences must meet requirements of specificity to a predetermined class of objects, as well as of consistency upon them.At the very beginning, we tried to determine such differences by using regular grammars working on edge languages, and as this model failed, we realized the edge information was not sufficient for our practical application. Then we built other segmentation primitives [Bla89] we called structures (Tightly Linked Structures) and, in the same time we defined a new kind of grammars, the T.L.S. grammars, in order to include these primitives inside the syntactic formalism. The matching of structures is performed in a single step by means of a recognizer based on T.L.S. grammars.In a few words, structures are defined in our application as particular links between confirmed segments. For example, given a military truck (figure 1), one can see it as a distorted rectangular