Construction of a multi-class discrimination matrix and systematic selection of areal texture parameters for quantitative surface and defect classification

Abstract

The surface of every manufactured component has a topography resulting from its fabrication route, which influences its final functionality (optical performance, wear resistance, aesthetic quality, etc.). While the quantitative characterization of the surface is essential for detecting process-induced variations or defects and predicting functional performance, the plethora of surface descriptors developed by the image processing and surface metrology communities makes the selection of optimal surface descriptors Edisonian in nature and highly dependent on experiential knowledge. This work proposes a systematic approach for selecting surface parameters that best distinguish between different surfaces as grouped by visual or process-related differences. Using a form of univariate analysis rooted in signal detection theory, the predictive capability of a discriminability value, d′, is demonstrated in the classification of mutually exclusive surface states. A “discrimination matrix” that offers a robust feature selection algorithm for multiclassification challenges is also introduced. The generality of the approach is demonstrated on the Northeastern University dataset consisting of intensity images from six different surface classes commonly found in hot-rolled steel strip operations. Using the outlined approach, it was found that only four surface descriptors used in conjunction with a simple decision tree classifier achieved a 95% classification accuracy. Surface descriptors used in the study were limited to those described within the ISO 25178-2 standard, while machine learning approaches were limited to a decision tree classifier. The reasoning for both is to maintain as much algorithm output interpretability as possible; the advantages of such is discussed from the perspective of the larger goal of linking surface texture to manufacturing processes and surface functionality fundamental mechanisms.