A hybrid descriptor for low-textural image stitching in real-time surface inspection systems

Surface inspection systems in the steel industry use multiple machine-vision (MV) cameras to inspect steel sheets for real-time quality control. Conventional approaches are classified into direct, deep-learning-based, and feature-based methodologies. Direct techniques perform poorly on parallax, while deep-learning-based algorithms require higher execution times and are ineffective for real-time applications. We propose a hybrid descriptor that uses defect detection to effectively stitch low-textural images captured by multiple cameras that are evaluated based on matching accuracy, execution time, and quality of stitched images and compared to popular feature-based image descriptor algorithms. Experimental results show that the proposed hybrid descriptor outperforms existing feature descriptors with 91% matching accuracy and an execution time of 49 milliseconds, producing a seamlessly stitched output.

Automated summary

Surface inspection systems in the steel industry use multiple machine-vision cameras for real-time quality control. Existing approaches, including direct and deep-learning-based techniques, face limitations in terms of parallax and real-time application effectiveness. We propose a hybrid descriptor that effectively stitches low-textural images captured by multiple cameras using defect detection. Experimental results demonstrate that our hybrid descriptor outperforms existing feature descriptors in terms of matching accuracy and execution time, producing a seamlessly stitched output.