Non-Contact Image-Based Roughness Estimation of Drilled Holes: Imaging Setup, Ground-Truth Dataset and Leakage-Free Evaluation of GLCM and Wavelet Features

Abstract

Surface roughness (Ra) of drilled hole walls is difficult to measure non-destructively because the recessed surface is inaccessible to a stylus unless the workpiece is sectioned. This study presents an imaging rig and a stylus ground-truth procedure for 150 holes (4–9 mm) drilled in Al-7075, and evaluates whether gray-level co-occurrence matrix (GLCM), first-order and wavelet features from grayscale hole-wall images, together with spindle speed and feed rate, can estimate Ra using support vector regression (SVR) with and without mutual information feature selection. Two evaluation schemes were compared: an image-level split of 900 augmented images, and hole-level cross-validation in which all images of a hole are kept together and tests use original images of unseen holes. The image-level split gave R2 = 0.80, whereas hole-level cross-validation gave R2 between −0.04 and 0.08 for all feature sets, showing that image-level splitting overstates performance for this kind of data. Flips left the features unchanged, whereas noise injection and 90° rotation altered them and their correlation with Ra. The work provides a reproducible acquisition and ground-truth procedure, a 150-hole dataset and an evaluation protocol for image-based roughness estimation of drilled holes, and identifies the data and imaging conditions needed for estimation on unseen holes.