Multi-line laser stripe centerline extraction methods for reflective metal surfaces using deep learning and Hessian matrix

Abstract Accurately extracting the center of laser stripes remains a challenge in industrial measurement, where precision and noise robustness must be balanced. On highly reflective metal surfaces, low surface roughness often leads to overexposed regions in laser stripe images. This results in uneven stripe width distribution and asylight intensity profiles, complicating center extraction. In this paper, we propose a novel method for laser stripe extraction from images of reflective metal surfaces, leveraging deep learning and multi-scale Gaussian kernels. First, an improved YOLOv8 network detects regions of interest (ROIs) containing laser lines and determines stripe cross-section widths. Next, a multi-scale Gaussian template, derived from the stripe width, is convolved with the ROI to construct a multi-scale Hessian matrix. Using the Hessian matrix, the normal direction of the laser line is calculated, along which a second-order Taylor expansion is performed to locate the subpixel center point. To enhance robustness, non-maximum suppression and a stripe energy distribution model are applied to eliminate noise points, optimizing the center coordinates. The entire algorithm is implemented with parallel acceleration, enabling high-precision and noise-robust subpixel center extraction. We validate the proposed method on both synthetic and real-world images. The experiments show that under different noise conditions, the deviation of the central point pixel remains within 0.09 pixels. Compared to traditional Steger’s method, our approach improves subpixel localization accuracy by up to 45.80%, effectively improving the sensor’s measurement performance.