Sistema de identificação da ordem de produção estampada em tarugos de aço

• Ano de defesa: 2005
• Autor(a) principal: Marcelo Cherem Ramalho
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Minas Gerais; UFMG
• Palavras-chave em Português: Redes neurais artificiais; Engenharia elétrica
• Link de acesso: http://hdl.handle.net/1843/HSAA-6M2PLW

This dissertation presents an overview of character recognition methods for application in a real problem of Gerdau Açominas. The problem consists of the identification of OP (Production Order) numbers imprint on steel objects (billets). The recognition will complete the automation of the inspection line. The parameters for the inspectionare defined by the OP number. After the inspection the billets are separated by client. The correct recognition prevent problems such as mixture of billets and the inspection of billets with wrong parameters. The study is carried through with the collected images, supplying requirements for the choice of the algorithms. The problem is divided in three stages: segmentation, feature extraction and classification. Methods are proposed in these areas, in two fronts. First, tests in images collected in laboratory are carried through, without conditioning the scale or rotation. The rotation is decided by the analysis of the texture present in the billets, diminishing the computational cost and supplying constantrotation for the stage of feature extraction. The method achieved a identification rate of 100% in images collected in laboratory. Most feature extraction methods are designed for solid binary images. A study is carried through on the different methods in literature. With the analysis of the collected images, a method for the extraction of the characterfrom the gray level image is proposed. In the stage of recognition, the method of tem- plate matching was tested, implemented by a normalized cross correlation. The method provided a recognition rate of 98% in images with better contrast and low deformation, which indicates that better conditioning of the illumination and scale are necessary. Thesecond front demonstrates the preliminary results in images collected in the inspection line. The identification of the rotation provided a rate of 94%. The images presents better conditioning of illumination and scale, supplying promising results for the implementation of the system.