Sistema de visão computacional para a classificação e medida da posição angular de objetos metálicos

Abstract

Computer vision techniques can be useful tools for production line automation in tasks such as the manufacture of electronic components, inspection of finished metal objects, production of printed circuit boards, among others. Often, there are production stages in metallurgical industries in which the identification of metallic objects, and their respective angular orientation, is needed either for analysis, separation or even only to classify these objects. The development of this project is motivated by the subsequent application in a real production line. The project use should reduce the line operating costs, reducing the error margin and increasing the production speed. The objective of this work is to develop a system capable of classifying metallic objects and measure their respective angle of rotation with respect to the x axis of the Cartesian plane. The objects are previously known by the system using machine learning techniques. The project development is divided into three steps. The first step is the study of techniques for image segmentation and implementation in MATLAB software. The second step is the study of techniques for image description and also implementation in MATLAB. Finally, in the third step a set of neural networks capable of classifying objects and measuring their angle is implemented using MATLAB. For each step in the project development a spiral development methodology is adopted. In each cycle of the methodology new features are aggregated to the system. In order to test the system a test environment is developed, where specifically selected objects are placed automatically, allowing the capture of images of the object in different angular positions. The system is divided into three modules: segmentation, description, and classification of objects. After capturing an image of the object being classified, the segmentation module selects the region of interest from the image. In the description module features are extracted from the region of interest, forming a descriptor that is provided to the third module. The third module then classifies the object and measures its angular orientation. Therefore, the result of the work is a system able to correctly classify 100% of the tested objects and measure the angle of these objects with accuracy of _4:5o in the worst case, using for that image processing and machine learning techniques.