Abnormal Wedge Bond Detection Using Convolutional Autoencoders in Industrial Vision Systems

Anomaly detection using machine vision images is of utmost importance for mission-critical tasks in industrial applications. Convolutional autoencoders have been widely adopted for unsupervised defect identification on vision data. This type of deep learning vision model is able to identify the anomaly pattern of target objects/areas without incurring the heavy workload on annotation/labelling. This paper investigates the challenging problem of anomaly detection in the quality control of wedge bond production. We report the progress of using SSIM (Structural Similarity)-based autoencoder in the abnormal wedge bond detection. The main idea is to leverage the advantages of both SSIM and autoencoder on identifying the abnormal features of wedge bond. SSIM analyzes the image difference in terms of luminance, contrast and structure. An autoencoder is trained to identify such visual differences in the prediction images. We conduct extensive experiments using images captured from real wedge production line for the anomaly detection. The evaluation results demonstrate the effectiveness and accuracy of the SSIM-based autoencoder.