Vitor Manuel Filipe

Abstract

Abstract
Quality automobile inspection is one of the critical application areas to achieve better quality at low cost and can be obtained with the advance computer vision technology. Whether for the quality inspection or the automatic assembly of automobile parts, automatic recognition of automobile parts plays an important role. In this article, vehicle parts are classified using deep neural network architecture designed based on ConvNet. The public dataset available in CompCars [1] were used to train and test a VGG16 deep learning architecture with a fully connected output layer of 8 neurons. The dataset has 20,439 RGB images of eight interior and exterior car parts taken from the front view. The dataset was first separated for training and testing purpose, and again training dataset was divided into training and validation purpose. The average accuracy of 93.75% and highest accuracy of 97.2% of individual parts recognition were obtained. The classification of car parts contributes to various applications, including car manufacturing, model verification, car inspection system, among others. © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2021.

Abstract
The automotive industry has an extremely high-quality product standard, not just for the security risks each faulty component can present, but the very brand image it must uphold at all times to stay competitive. In this paper, a prototype model is proposed for smart quality inspection using machine vision. The engine labels are detected using Faster-RCNN and YOLOv3 object detection algorithms. All the experiments were carried out using a custom dataset collected at an automotive assembly plant. Eight engine labels of two brands (Citroën and Peugeot) and more than ten models were detected. The results were evaluated using the metrics Intersection of Union (IoU), mean of Average Precision (mAP), Confusion Matrix, Precision and Recall. The results were validated in three folds. The models were trained using a custom dataset containing images and annotation files collected and prepared manually. Data Augmentation techniques were applied to increase the image diversity. The result without data augmentation was 92.5%, and with it the value was up-to 100%. Faster-RCNN has more accurate results compared to YOLOv3. © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2021.

Abstract

Abstract