Automatic segmentation of metal surface defects with noise label

Abstract Defect detection on metal surfaces is crucial for quality control and precise maintenance in various industrial scenarios. Nevertheless, metal surface defects demonstrate various variations and intricacies, making identifying high-precision metal surface defects a formidable task. With the development of artificial intelligence technology, metal surface defect detection technology has ushered in a huge breakthrough. Compared to manual inspection, metal surface defect detection methods based on computer vision and artificial intelligence technology have apparent advantages in detection efficiency and accuracy, becoming the mainstream means of detecting surface defects in metal parts. While deep learning techniques have shown impressive results in defect detection, the performance is highly susceptible to the influence of annotation quality. Using unreliable annotations to train deep learning models fails to fully utilize the superior performance of deep learning techniques and may even cause model performance degradation, seriously affecting predictive performance. To address these challenges, we propose a novel segmentation-based metal surface defect detection model with unreliable annotations or noise labels. First, we train a semantic segmentation network using pre-existing labels, employing an encoder-decoder architecture. Then, the trained model is employed to rectify the unreliable annotations and enhance the network's resilience through retraining. Our method is validated through adequate experimentation on two publicly accessible datasets, which show superior performance in metal surface defect segmentation. The studies demonstrate that even if there is noise in data annotation, our method can effectively suppress the predictive performance degradation of deep learning models.