Towards low-cost surface inspection using artificial vision in practical industrial environments

Surface inspection in the steel industry plays a crucial role in quality control, driving significant research efforts, particularly with the rise of deep learning models. However, many existing works focus on minor model variations that offer limited practical advantages and fail to address critical challenges such as lighting conditions and dataset curation. This paper presents a high-performance surface inspection system that integrates deep learning with real-time parallel processing to enhance defect detection in industrial settings. The proposed system employs a modular and scalable architecture based on a publish/subscribe messaging pattern, ensuring efficient data flow and seamless integration of components. It incorporates active learning methodologies and advanced image processing techniques, including color correction, light compensation, and contrast enhancement. Designed for industrial deployment, the system balances high performance with cost-effective installation and minimal maintenance by utilizing a single areascan camera and AC illumination. The optimized positioning of the camera and lighting reduces surface reflections and motion blur, improving defect detection accuracy. The system achieves high-speed processing by parallelizing preprocessing and post-processing tasks, effectively minimizing computational overhead and enabling continuous real-time monitoring. A built-in performance monitoring module provides valuable insights for system optimization and real-time assessment. Extensive testing in a steel pickling line demonstrates the system’s robustness and effectiveness in real-world industrial environments, offering a reliable and scalable solution for automated surface inspection.