Ensembling Deep Learning Models for Metal Surface Defect Classification

Abstract Defect detection on metal surfaces is crucial in manufacturing because it helps enable industrial automation. Manual inspection and simple image processing techniques, which are commonly used for defect detection, are inefficient and have low accuracy. This paper presents an ensemble based learning model to analyze and classify metal surface images with defects. By ensembling two Convolutional Neural Networks (CNNs) trained using two different pre-processing techniques, the proposed method can automatically categorize six types of defects (crazing, inclusion, patches, pitted surface, rolled-in scale, and scratches). The two CNNs are ensembled using stacking where the meta learner is a Support Vector Machine (SVM). In the experiments, a dataset containing 300 images of each type of metal surface defect (1,800 images in total) is used to train and test the model. The results demonstrate that the proposed method surpasses traditional methods. The accuracy of the ensemble after 10-fold cross-validation reaches 96.7% on the Northeastern University Classification of Surface Defects dataset, which is higher than the accuracy of the individual CNNs and competitive with the current state-of-the-art. On a dataset with non-defect images included, the accuracy reaches 96.3%. The main contribution of this study is to provide deep learning based methods that are accurate surface defect identification.