Research on a Surface Roughness Measurement Under ResNet-Based Roughness Classification and Light-Section With Seam-Driven Image Stitching (RCLS)

Abstract

With the development of optics, light section method has become a feasible measurement for surface roughness, while the short sampling length is negative to the accuracy. To overcome this defect, this article proposed a measurement under ResNet-based roughness classification and light section with seam-driven image stitching (RCLS). First, the images were classified with ResNet neural network, then stitched and enhanced by scale invariant feature transform (SIFT) and optimized random sample consensus (RANSAC) algorithm for the best visual effect. After this, images were processed by Nobuyuki Otsu method and Freeman chain code tracking algorithm. Least square was also adopted to calculate the optical band edge curve and contour midline. Finally, the roughness contour arithmetic mean deviation model was established to evaluate the surface roughness. The experiments were conducted with vertical milled, planned, and turned samples that self-machined. The light section method had a reduction of 2.75% on the mean relative error compared to stylus and RCLS could further reduce the mean relative error by 1.42%, especially in planned sample. The RCLS could achieve a more accurate surface roughness by overcoming the disadvantages of small sample length and low precision of the light section method, and is more convenient than stylus.

Keywords

• Elongated sampling length
• image stitching
• light section
• seam driven
• surface roughness measurement