Anomaly detection for machined surfaces of castings using CBiGAN

Abstract

For image anomaly detection, various DCGAN-based methods have been proposed with the improvement of image processing performance by machine learning. CBiGAN is a method that achieves high performance. However, when CBiGAN is applied to the machined surface of a casting, several issues arise in normal regions. First, it is difficult to generate samples with biased image brightness. Second, DCGAN tends to generate average images, which makes it difficult to generate random patterns with high contrast. Third, the accumulation of small anomaly scores for normal regions increases the anomaly score of the entire image, making classification difficult. In this paper, we propose methods to improve these issues. We introduce normalization of image brightness, masking to confine the target processing area, and parameters to differentiate between high and low anomaly scores. The proposed methods improved the performance of anomaly detection by separating the distribution of anomaly scores for normal and anomaly samples.