Musings on Symbol Recognition

Symbol recognition is a field within graphics recognition to which a lot of efforts have already been devoted. However, a document analysis expert who is more familiar with OCR might rightfully wonder what exactly we call a symbol and how symbol recognition differs from basic character recognition. Our feeling is that the problem is very different because of the much higher number and variety of symbols to be recognized. Except in strongly context-dependent applications, it is impossible to provide a database of all possible symbols. It is also in many cases impossible to assume that symbol recognition can be performed on clearly segmented instances of symbols, as symbols are very often connected to other graphics and/or associated with text. The well-known paradox therefore appears: in order to correctly recognize the symbols, we should be able to segment the input data, but in order to correctly segment them, we need the symbols to be recognized! This in turn means that it is usually not possible to perform symbol recognition by simply assuming that a reliable segmentation process is available, that the symbols have been clearly extracted, normalized, etc. and that a vector of general-purpose features can be computed on these symbols to be recognized, in such a way that the vector can be classified by some appropriate statistical pattern recognition method. The most common approach in symbol recognition relies on structural methods able to capture the spatial and topological relationships between graphical primitives; these methods are sometimes complemented by a classification step once the candidate symbol has been segmented or spotted. This paper does not pretend to be yet another survey on symbol recognition methods, as several excellent surveys already exist [4, 6, 23]. We will rather try to take a step back, look at the main efforts done in that area throughout the years and propose some interesting directions to investigate.