This paper presents a general robust OCR system designed for practical use and suited to unconstrained gray-level images grabbed from a CCD camera. The system works with minimum assumptions on font , text location, size, color and the background scene. To avoid the threshold problems with text segmentation and feature extraction, which depend on character binarization, we propose to process the full gray level image directly. The text blocks localization in complex scenes using a specific filter which enhances any text from the backgound without binarization. A special stage is designed to separate characters, even touched, by using gray-level information. We also extract gray-level features which make the algorithm more reliable, in particular under poor printing conditions or bad contrast digitization.