A review of deep learning-based surface defect detection for castings

As a core technology in the high-end equipment manufacturing industry, casting requires crucial surface defect detection of products for quality control. Traditional nondestructive testing methods relying on manual experience present limitations including low efficiency, strong subjectivity, and poor adaptability, failing to meet modern industrial automation inspection requirements. Deep learning technology addresses the inefficiency of manually designed features in conventional approaches through convolutional neural networks and other models that automatically extract deep image features, enabling end-to-end defect detection. Current research primarily focuses on two-stage and single-stage object detection algorithms - the former emphasizes accuracy but suffers from computational complexity, while the latter prioritizes speed at the expense of small target detection capability. Domestic and international studies continue to optimize detection performance through network structure improvements, lightweight design implementation, and efficient loss function integration. Future research should enhance model generalization in complex scenarios, balance detection accuracy with operational speed, explore multi-modal data fusion and self-supervised learning technologies, thereby advancing casting defect detection systems toward intelligent and efficient development to provide technical support for industrial quality management.