Autoencoder-based anomaly detection for surface defect inspection

Abstract

In this paper, the unsupervised autoencoder learning for automated defect detection in manufacturing is evaluated, where only the defect-free samples are required for the model training. The loss function of a Convolutional Autoencoder (CAE) model only aims at minimizing the reconstruction errors, and makes the representative features widely spread. The proposed CAE in this study incorporates a regularization that improves the feature distribution of defect-free samples within a tight range. It makes the representative feature vectors of all training samples as close as possible to the mean feature vector so that a defect sample in the evaluation stage can generate a distinct distance from the trained center of defect-free samples. The proposed CAE model with regularizations has been tested on a variety of material surfaces, including textural and patterned surfaces in images. The experimental results reveal that the proposed CAE with regularizations significantly outperforms the conventional CAE for defect detection applications in the industry.

Introduction

Machine vision is an effective non-contact technology for automated defect inspection in the manufacturing process. Most of the traditional machine vision techniques are based on texture analysis. A set of discriminative features are extracted from the spatial or the spectral domain of the test image. A high-level multiple dimensional classifier such as Support Vector Machine (SVM) or Random Forest is then applied to identify defect samples. The success of the classification highly relies on the human experts to extract and select representative features based on the local gray-level (or color) and structure variations of a defect in the test image.

In the manufacturing environment, it is quite easy to collect normal samples as many as required. However, it is difficult to collect a sufficient number of defective samples in a short period of time to train robust classification models for defect detection. The machine vision methods currently available need handcrafted features based on the characteristics of individual defect types of a specific product, where the defect samples may not be sufficient for the analysis.

In this paper, the deep learning technique is explored to tackle the defect detection task without defect samples for the training. The proposed method is image-wise defect detection, i.e. it classifies a test image as defective or defect-free. It is not used for pixel-wise defect segmentation. The unsupervised convolutional autoencoder (CAE) is applied to extract the representative features that can well describe the distribution from a set of normal samples. The loss function of the conventional CAE measures only the reconstruction errors. It could make the extracted feature values widely spread in the high-dimensional variable space. When the trained CAE is used for anomaly detection, the encoded features of a defect sample image may thus fall within the range of the normal samples’ variable space. It could make the extracted features indistinguishable between normality and abnormality. A regularization penalty is thus included in the original CAE loss function to limit the spread of the learned feature values for normal training samples. It is expected that the distances of feature vectors between the normal samples are close to each other and the unseen defect samples yield sufficiently large distances from the normal samples. The proposed CAE-based model, denoted by $\lambda$-CAE, is tested on various material surfaces for defect detection, including textural surfaces and patterned surfaces. The proposed method is also compared to autoencoder-variant models based on encoded features and image reconstruction error.

This paper is organized as follows. Section 2 reviews the related work on defect detection with traditional machine vision and deep learning techniques. Anomaly detection with autoencoders for various applications is also discussed. Section 3 presents the original CAE and the regularized $\lambda$-CAE models for defect detection. Section 4 discusses the experimental results on various material surfaces. The paper is concluded in Section 5.