Research on binocular 3D measurement method for notch damage in aero-engine blades based on geometry-aware keypoint localization

Abstract

Accurate quantification of notch damage from in-service aero-engine blade borescope images is vital for maintenance decision-making. Traditional manual point-selection and heatmap-regression-based keypoint localization methods often fail to deliver stable industrial-level measurement accuracy under complex imaging conditions including strong reflections, low contrast and cluttered backgrounds. To tackle this problem, this paper proposes a full-process intelligent measurement framework for complex industrial scenarios to realize non-contact three-dimensional (3D) measurement of blade notch length and depth. First, a lightweight scale-guided reflection analysis detection network is constructed to detect damaged region of interests in binocular images and suppress metal highlight interference. Second, a geometry-aware SimCC network with decoupled classification heads and geometric prior loss is presented for pixel-level notch keypoint localization. Finally, semantic matching and epipolar constraints are combined to build robust binocular correspondences. 3D keypoint coordinates are recovered via triangulation, and notch physical dimensions are computed using spatial geometric definitions. Experiments show average absolute errors of 0.685 mm (length) and 0.873 mm (depth), with an average keypoint localization error of 2.86 pixels. Maintaining a lightweight architecture and low computational cost, the proposed method achieves sub-millimeter-to-millimeter-level measurement precision, satisfying industrial borescope preliminary screening tolerances and offering a feasible solution for quantitative assessment of aero-engine blade notch damage.