According to characteristics of the pressed character on cylindrical surface with the same color as the background, the curved surface and the poor image quality, a scaling transformation method for local grey interval maximization is adopted to highlight the character region. Meanwhile a character segmentation method selecting character region before morphological optimization was proposed to achieve the optimization of character outline, and overcome the shortage of character adhesion in interference region caused by the traditional method using morphological optimization directly. Finally, the training file was created to train the Back-propagation (BP) neural network, and recognition results were displayed. The experimental results show that the recognition algorithm of the pressed character has a good ability of robustness and high accuracy, entirely meeting the industrial requirements.