Detecção de defeitos em tecidos através de redes neurais convolucionais

• Autor(a) principal: Souza, Lucas Candiani
• Data de Publicação: 2022
• Tipo de documento: Trabalho de conclusão de curso
• Idioma: por
• Título da fonte: Repositório Institucional da UFSCAR
• Texto Completo: https://repositorio.ufscar.br/handle/ufscar/16102
• topic: Reconhecimento de textura; Aprendizado de máquina; Aprendizado por transferência; Texture recognition; Machine learning; Transfer learning; CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::SISTEMAS DE COMPUTACAO; ENGENHARIAS::ENGENHARIA ELETRICA::ELETRONICA INDUSTRIAL, SISTEMAS E CONTROLES ELETRONICOS

This work aims to apply transfer learning to a detector based on convolutional neural networks to identify defects and patterns through tissue images. With Industry 4.0, intelligent systems are increasingly being applied in an integrated way in production; therefore, the objective of this work is to develop a model to perform the automatic identification of fabric patterns and defects, since an effective detector applied in a production line can reduce costs and provide data on production in general. The dataset used was the ZJU-Leaper, the backbone used was from RetinanetR101 and for code execution several Python libraries were used, such as Pytorch, OpenCV and numpy, in addition to the Detectron2 API. The mAP (mean average precision) of the model developed, calculated for all classes (6 patterns and “defect”) in the validation set, is 92.4%.