Análise comparativa de redes neurais convolucionais para classificação de fibras ópticas.

Abstract

This paper utilizes deep learning-based algorithms to solve an industrial field problem, which is the classification of fiber optics’ quality. One of the motivations in this work is the lack of related work in the defined segment. There is currently too little material available concerning computer aided vision to quicken this kind of process. The evaluation method was through a training of two Convolutional Neural Networks, in order to compare their parameters (such as training and validation accuracies, training and validation losses, training time and predictions on a testing dataset). Chosen CNNs were VGG-16 and ResNet-18, due mainly to their usual utilization in computer vision tasks. The utilized dataset had 440 fiber optics’ images in different states of preservation (new and clean, new and dirty and damaged) and in five different colors (red, blue, green, yellow and white). To those fibers it was assigned one of two different classes (“good” or “bad”). The dataset is a result from a data augmentation algorithm plus 27 original images. With those parameters, both neural networks have achieved expressive accuracy results, with VGG-16 performing a validation accuracy of 93,18% and ResNet-18 going as high as 96,21% on the test dataset.