Edge AI-based Self-Supervised Surface Defect Localization for Industrial Visual Inspection

Abstract

The industrial visual inspection problem is to locate the surface defects with accurate and limited number of anomaly labels, limited resources and tight line speed requirements. Traditional supervised segmentation approaches demand high quality pixel masks, and many unsupervised or reconstruction-based anomaly detectors are either synthetic anomaly mismatch, lack of boundary accuracy, large memory banks, or do not have real-time inference. In this paper, we present a framework for surface-defect localization using self-supervised representations of normal surface images and sparse annotations to adapt to real defect images. The framework consists of a light-weight self-supervised encoder, realistic pseudo-defect synthesis, sparse point/box/scribble guidance, compact normal-prototype memory, multi-scale residual scoring and edge-aware compression. Mathematical formulations are given for the self-supervised objective, the sparse localization prior, the anomaly score, and the edge deployment optimization. A benchmark simulated environment is mentioned to demonstrate the potential localization quality, latency, throughput and ablation for the edge industrial inspection scenario. The proposed design decreases the cost of annotation, enhances interpretability using heatmap overlays and enables deployment on small edge devices for real-time OK/NG decisions.