Identifying Pristine and Processed Animal Fibers using Machine Learning

Animal fiber identification is a crucial aspect of fabric production, as specialty fibers such as cashmere are often targeted by adulteration attempts. Since animal fiber identification is difficult, it is currently performed by human experts using Scanning Electron Microscopy (SEM). Many algorithms have been proposed to tackle the automated identification of animal and specialty fibers in SEM images. While ever-increasing classification performance is reported, the adulteration resistance of the proposed methods has not yet been evaluated. In our work, we perform such an evaluation for the first time. Lacking known ground truth adulterations, we construct a dataset containing specialty as well as conventional animal fibers in a chemically treated and untreated setting, where treated and untreated state differ slightly. We subsequently benchmark the ability of proposed state-of-the-art methods to correctly identify animal fibers including treatment status as a surrogate for adulteration resistance. Our results reveal that not all methods are equally capable at distinguishing treated and untreated fibers. Therefore, future research on animal fiber identification should additionally focus on adulteration resistance.