Human perception for print defect detection

Human perception plays a critical role in industrial quality control, where defect tolerability is often subjective and inconsistently assessed. We present a perception-centered evaluation framework for anomaly detection (AD) systems, emphasizing how comparisons between potentially defective images and their reconstructed, defect-free counterparts can be aligned with expert visual judgment. To support this, we synthetically generate realistic packaging foil designs and simulate two common defect types: ink spots and missing colour. Expert raters provide binary acceptability labels, capturing the variability in human defect tolerance. Central to our approach is a perceptual assessment module that quantifies the severity of visual anomalies by comparing input images with their corresponding clean reconstructions. Trained on synthetic perceptual scores, this module outperforms traditional full-reference image quality assessment (FRIQA) methods. With minimal fine-tuning on expert judgments, we achieve inter-rater agreement on par with human experts, as measured by Krippendorff's Alpha. We also contribute a new copyright-free dataset and evaluation protocol to facilitate future work in perception-aligned quality control.