Surface Defect Detection of Casting with Machined Surfaces Based on Natural Artificial Defects

Defect detection of the casting’s surface is of great importance for its quality inspection. There are a series of challenges of clear imaging and weak contrast between defects and surroundings for casting surface defects. In this paper, we propose a feasible scheme that an imaging system could combine physical exposure compensation with algorithm compensation to deal with the difficulty of large differences in light reflectivity and scattering between casting and machining surfaces. Then, we present the image preprocessing algorithm which is based on the Hough line transform and Canny edge detection. Furthermore, for alleviating the shortage of datasets in this field, we provide a representative Casting Surface Defect(CSD) detection dataset containing 12647 high-resolution gray images of six surfaces. Besides, we present pixel-precise ground truth zones for all defect samples. We also present a comprehensive evaluation of state-of-the-art detection methods. Natural Artificial defects (NAD) are introduced to improve Casting Quality Inspection Efficiency. This results in a wide range of composite defects more similar to real defects than ever. Experiments show that a model trained to classify synthetic anomalies from normal samples generalizes well on real-world manufacturing defects. Our method outperforms competitive approaches on six surface datasets of the casting. Keywords: defect detection; localization; self-supervised learning; Natural Artificial Defects