Syntactic approach to image analysis (abstract only)

Abstract

During the past several years, syntactic approach [1,2] has attracted growing attention as promising avenues of approach in image analysis. The object of image analysis is to extract as much information as possible from a given image or a set of images. In this abstract, we will focus our attention on the use of semantic information and grammatical inference.In an attributed grammar, there are still a set of nonterminals, a set of terminals and a start symbol just as in conventional grammars. The productions are different. Each semantic rule. Two kinds of attributes are included in the semantic rules: inherited attributes and synthesized attributes. One example of the attributes is the length of a specific line segment used as a primitive. All the attributes identified for a pattern are expressed in a “total attribute vector”.Instead of using attributes, stochastic grammars associate with each production a probability. That means, one sub-pattern may generate one subpattern with some probability, and another with a different probability. A string may have two or more possible parses. In this case of ambiguity, the probabilities associated with the several possible productions are compared to determine the best fit one. Probabilities are multiplied in multiple steps of stochastic derivations.Besides these, fuzzy languages[3-6] have also been introduced into pattern recognition. By using similarity measures as membership functions, this approach describes patterns in a more understandable way than stochastic grammars. Moreover, fuzzy languages make use of individual characteristics of a class of patterns rather than collective characteristics as in stochastic languages, and therefore it is probably easier to develop grammars than stochastic languages. Yet a lot of work still need to be done in order to develop sufficient theories in this field for practical uses.An appropriate grammar is the core of any type of syntactic pattern recognition process. Grammars may be established by inferring from a priori knowledge about the objects or scenes to be recognized. Another way to establish a pattern grammar is by direct inference from some sample input patterns.Once a grammar is derived from some sample input patterns, other patterns similar to them or belonging to the same class can be parsed according to the grammar. Therefore grammatical inference enables a system to learn most information from an input pattern, and, furthermore, to apply the obtained knowledge to future recognition processes. It seems to be the ultimate aim of image analysis.Inference can be supervised or unsupervised. In supervised inference, a “teacher” who is able to discriminate valid and invalid strings helps in reducing the length of sentences or inserting substrings until some iterative regularity is detected. In unsupervised inference, no prior knowledge about the grammar is assumed.The difficulty of inference is proportional to the complexity of the grammar, and the inference problem does not have a unique solution unless some additional constraints are placed upon the grammars. Some theoretical algorithms have been developed for inferencing regular (finite-state) grammars, but they still have severe limitations for practical use because of large amount computation due to the combinatorial effect.Context-free grammars are even harder to deal with since many decidable properties about regular grammars are undecidable for context-free grammars, such as the equivalency of two contex-free grammars. Therefore, inference algorithms have been developed only for some specific types of context-free grammars and most of them rely on heuristic methods.Syntactic approach to image analysis may be applied to many areas including space object surveillance and identification [7].