Research On Symbol Recognition And Its Application In Traffic Sign Recognition

Shape retrieval and symbol recognition play an important role in automatic analysis of documents, including music notations, mathematical expressions, engineering drawings and historical files. An ideal symbol recognizer should be domain-independent, scalable, computationally efficient, robust in the presence of noise and distortion, and rotation and scale-invariant. So far, the effort is ongoing to look for a method that can promise a satisfactory overall performance, but it is hard to find an algorithm that performs well in all the aspects.This paper focuses on the symbol recognition problem in the case of noise, rotation, translation and other degradation. We present a novel descriptor named Multiple References Histogram Matrix (MRHM) to retrieve shapes. MRHM first generates a set of uniform grids for each shape and then enhances the shape contours and simultaneously filters noises in each grid by line Hough transformations and curve-fitting transformations. Next, MRHM selects one reference point for each grid and calculates its local distributions. Finally MRHM integrates all the local distributions into a global distribution matrix for shape matching.Our method has the following advantages against the existing methods:1) we can first divide the shape into a number of grids and then use Hough transformations in each local grid to track line or short curve segments probably existing on the shape contour. Any type of curves can be tracked as long as the subdivided grid is small enough. Once the straight line segments and the curves fitted by short line segments are recognized, the shape contour can be well enhanced and the background noises can be effectively filtered simultaneously;2) the method is applicable to various symbols in the different areas; and3) we explore the problem of traffic sign recognition in the natural seen and then design a traffic sign recognition system based on the geometric constraint distribution, the experiment shows that the system is applicable to traffic sign recognition.