Eventos Anais de eventos
COBEM 2023
27th International Congress of Mechanical Engineering
Mechanical components recognition through computer vision usage
Submission Author:
Leonel Fernandes Balbino , PR
Co-Authors:
Leonel Fernandes Balbino, Carlos Alberto Fiakofski Cadamuro, Giuliana Sardi Venter
Presenter: Giuliana Sardi Venter
doi://10.26678/ABCM.COBEM2023.COB2023-1847
Abstract
Artificial intelligence has been a field of great relevance in several applications across all industry sectors. The advances achieved in terms of computational power, cheapness and availability of infrastructure have provided, in recent years, a complexification of information systems and possibilities for implementing resources that were previously inaccessible. Still at the beginning of the last decade, in 2012, convolutional neural networks became the preferred tool in image recognition and segmentation models when in a global level competition in classification of a dataset with 1000 classes, a group of researchers from the University of Toronto obtained the best position using such an architecture. Broad usages of computer vision comprise face recognition on security services like smartphone facial biometrics and document authenticity verification, anomaly detection on image diagnosis through CT scan and X-ray, detection of crop diseases, document digitalization by "reading" the pages and converting it to digital content and many other tasks. Specifically in Engineering and Manufacturing fields, there already are applications of computer vision systems on autonomous vehicles and production quality control, for example. The present work aims as its global objective to develop a classifier of mechanical parts through images in video stream. As its four specific milestones there are the obtaining of an heterogeneous dataset based on a set of mechanical components 3D models gathered from engineering supplier catalogs, the training of three largely known architectures with transfer learning and comparison of the best result obtained from this set with the result of a network fully trained over the dataset for feature extraction, comparison of the result obtained from the standard dataset with the result obtained on preprocessed images with different textures and, finally, the deployment of the model to be tested over real time video stream for image segmentation. The work towards achieving these four milestones consists in defining a sufficient number of views and their orientation for the 65 thousand 3D models available, selecting the architectures to be used, applying data augmentation techniques on the image dataset to minimize overfitting, comparing results and deploying the selected model served as a web application to be tested using real video streams captured in controlled scenarios simulating production environment.
Keywords
Convolutional neural networks, Computer Vision, image segmentation
