Title: Systematic Data Acquisition for Advancing AI-Driven Quality Control in Carpet Production

Abstract

Visual inspection continues to serve as the standard for quality assurance in both woven and tufted carpet manufacturing. However, this manual approach is plagued by inconsistency, subjectivity, and sluggishness, particularly when operating at the high speeds and widths characteristic of contemporary looms. This paper introduces a design framework for an inline machine-vision system with a dual mandate: to perform real-time inspection of the carpet web and, equally critical, to systematically gather and annotate images of defect patterns. This data infrastructure enables the continuous training of increasingly sophisticated quality-control models throughout the operational lifespan of the installation.

The proposed solution is rooted in a specific industrial context: a Six Sigma (DMAIC) initiative at a woven-carpet facility. This project emerged in anticipation of a production bottleneck triggered by the addition of new weaving machines, compounded by a high baseline defect rate and considerable financial risk tied to quality failures. We detail an imaging subsystem utilizing synchronized line-scan cameras equipped with combined bright-field and grazing illumination. Furthermore, we calculate the resolution and throughput specifications necessary to identify fine structural defects across multi-meter wide webs and establish a defect taxonomy specific to the carpet industry.

Our approach outlines a phased modeling strategy. It initiates with unsupervised anomaly detection models trained exclusively on defect-free samples, adhering to the methodology demonstrated in the carpet category of the MVTec Anomaly Detection benchmark. As the system evolves, it incorporates a human-in-the-loop annotation flywheel, eventually transitioning to supervised detection and segmentation models. Finally, we link detection performance directly to DMAIC goals, illustrating how minimizing escaped defects enhances overall process quality and increases process sigma levels. This work provides a comprehensive, deployable blueprint that elevates data collection from a secondary consideration to a primary engineering priority.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC