Técnicas de visão computacional para classificação de peças

Abstract

Computer vision is an area of artificial intelligence that has a great impact on various sectors of society. The ability of the machine to see objects of interest in an image and produce a classification or detection response is very important in the context of automation. The computer vision techniques allow, through image processing steps and the use of classifiers, to provide answers to several problems that arise. The objective of this work is to analyze the responsiveness of a specific classifier, the K-nearest neighbors (KNN), for a group classification problem and analyze its performance through hit rate and accuracy parameters. First, four groups of parts were created in a three-dimension (3D) design software and cut on a laser cutting machine, then pictures were taken of each of the parts individually. Later, using the OpenCV libraries, we could carry out the image processing, such as changing the red-green-blue (RGB) pattern to grayscale, binarization, noise correction, and obtaining invariant moments; and in the Scikit-Learn library, the classifier was trained and tested. It was possible to conclude that the classifier, in tests 1 and 2, was able to classify properly the parts groups having recall and precision above 65 and 80%, while, in test 3, it had a recall of 42% and a precision of 55%, showing that, with more analyzed classes, the classifier decreases its efficiency.