Machine Learning for Automatic Classification of defective Automobile Parts Images

Abstract

In this work, we use deep learning to automatically classify images of automotive parts which could have or have no defects. Particularly, we analyzed an automotive part which is indeed composed of two different pieces welded in 8 specific points. The absence of one or more welding points on the piece is considered a defective part.  We generate a database with 51 images of several parts containing up to eight welding points. We used a SegNet convolutional neural network with 14 layers. The neural network was trained using hand-labelled defect images. The results of the training stage are the specific weights corresponding to the characteristics of the images. Our results show that our method can locate accurately the welding points.