Cognition-Inspired Deep Learning Pipeline for Robust Surface Defect Inspection in Manufacturing

View at publisher Summarise with VIRAT
Authors: L. M. A. P. Cabral, Karishani Malhari Bamunawita
Cited by: 0
References: 35

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Surface defect inspection is a pivotal component of Quality Control (QC) in contemporary manufacturing and it is essential for maintaining production stability and product integrity. Conventional Machine Vision (MV) systems, though capable of higher throughput, still lack the cognitive adaptability required to manage environmental noise, non-uniform backgrounds, and complex non-planar object geometries. This paper presents a deep-learning inspection pipeline that explicitly models key aspects of human visual cognition: separating object from background, accounting for 3D shape and suppressing known non-defective regions. The architecture combines SAM v2 for object segmentation, Depth Anything v2 for monocular depth estimation, Grounding DINO plus SAM v2 for masking foreign objects, U-Flow-based anomaly detection and a custom ResNet-inspired classifier, integrated into an industrial imageacquisition and visualization system. Experiments on representative categories of the MVTec AD benchmark and a custom fruit dataset show robust anomaly localisation under illumination, pose and background changes; foreign-object segmentation achieves mean IoU$\approx \mathbf{0. 8 8}-\mathbf{0. 9 0}$and Dice$\approx \mathbf{0. 9 1}-\mathbf{0. 9 2}$, while U-Flow offers the best overall balance of AUPR and IoU across most tested categories. These results indicate that a cognition-inspired, proposed modular design is a promising direction for deployable automated optical inspection (AOI) in surface defect inspection.

Fields of Study & Subjects

Industrial Vision Systems and Defect Detection Infrastructure Maintenance and Monitoring Non-Destructive Testing Techniques

Publication details

Journal / source
International Conference on Advanced Research in Computing (ICARC)
Pages
1–6
Publication date
18 February 2026
Language
EN