Weld Cross-Section Profile Fitting and Geometric Dimension Measurement Method Based on Machine Vision

A visual measurement method based on a key point detection network is proposed for the difficulty of fitting cross-sectional profile ultra-narrow gap welds and low efficiency accuracy manual geometric parameters. First, HRnet (High-Resolution Net) algorithm was used to train feature model, 18 points in “measurement unit” were extracted. Secondly, coordinates are transformed from image coordinate system weld system, profiles fitted by least squares method. Finally, calibrated with coplanar linear calibration perform pixel distance actual conversion quantitative dimensions. The experimental results show that localization 95.6%, mean value coefficient determination R-square curve greater than 94%, absolute error between 0.06 0.15 mm, relative 1.27% 3.12%. more reliable, significantly improved compared measurement.