A robust recognition approach for low resolution vehicle license characters

The quality and size of a character image are the key factors to the recognition ratio. Unfortunately the character images of vehicle licenses are declining, small and noisy. This paper presents a high-accuracy and noise-resistance vehicle license recognition approach. The images are firstly corrected by a self-prediction correction algorithm and de-noised by, several smoothing algorithms for different types of noises, e.g. the modified mean filtering and a run-number smoothing method, which improves, the quality of input character images notably. Then a multi-layer classifier combining "main features" with "secondary features" is adopted to recognise the characters, The main features are extracted from the weighting Fourier transform coefficients, which are translation -invariant. As the result, this method achieves high recognition accuracy.