Inspeção e classificação de picos em peças estampadas de metal utilizando rede neural convolucional

Abstract

In this dissertation we present a framework for the automatic detection of defects in stamped metal parts used in automotive industries. Specifically, we address a production defect known as “imprint” --- protrusions in the direction of stamping mostly formed by the presence of dirt and debris in the stamping material.The fast detection of imprints in the automotive production line is essential to reduce the costs associated with these defects, which depending on their nature and extent, may cause a significant disposal of parts and repairs. Currently the process of inspecting the stamped parts is performed visually by quality inspectors, being a slow process and prone to failures. To assist in this task a study was carried out with two convolutional neural network architectures widely used for image classification, VGG16 and Resnet. The networks classify surfaces into three classes: surfaces without defect, with moderate imprints and severe imprints. In the first experiment the VGG16 network achieved 99.4% accuracy. A classification by using a support vector machine with Gabor filters achieved 80.7% accuracy in the same experiment. In a second experiment, using a more challenging database, we compared variations of the network architectures VGG16 and Resnet. The highest accuracy achieved was of 90.1% using the Resnet architecture. Finally, we performed a test outside a controlled environment in the painting stage, where the Resnet network was able to correctly identify the position of a severe imprint in the vehicle part, demonstrating the versatility of the proposed technique.