Automatic Selection of White Paint Types for Automotive Industry

Abstract

In this work, we design an optical experimental setup and a data augmentation methodology to generate 384 images of automotive parts, painted with two different types of white paint. We also perform the automatic classification of the database considering two different classes. To obtain the highest precision, we have used two different classification scenarios, 3 algorithms, and 4 metrics. Also, we assume that the results can be improved by extracting the image characteristics using the convolutional neural network ResNet50 and using them as input. Our results show that an error-free classification can be obtained independently of the scenario or classifier. We obtained 100 % in each of the 4 metrics in the six studied variations. Therefore, the machine time is the parameter we can use to select the optimal classifier, where the classifier Multinomial Naïve-Bayes under the Training and Test Set scenario was the fastest algorithm with 65 s, and the classifier Machine Vector Support under the Cross-Validation scenario, the slowest with 79.12 s.