A lightweight edge-vision measurement system for real-time inspection of concealed industrial product defects

Abstract

Automated visual inspection is crucial for industrial quality control, yet existing technologies struggle with highly concealed defects—characterized by blurred boundaries and heavy background noise—and extreme scale variations. To address these challenges, we propose the multi-scale dynamic edge enhancement network (MDEEN), an efficient edge-vision measurement system designed for real-time, automated defect inspection. MDEEN incorporates three key components: an edge feature enhancement module utilizing frequency-domain analysis to amplify weak signals for sharper boundary perception; a dual-axis self-calibrating module that calibrates spatial attention to suppress homogeneous texture noise; and a multi-kernel receptive field fusion module to aggregate features across extreme scale variations. Experimental results on NEU-DET steel, tire, and aluminum datasets demonstrate mAP@0.5 scores of 99.44%, 95.61%, and 98.33%, respectively. Under the stringent mAP@0.5:0.95 metric on NEU-DET, MDEEN achieves 82.61%. With a compact architecture of only 5.388 M parameters, the system achieves a real-time inference speed of 80.6 FPS on the NVIDIA Jetson Orin Nano, demonstrating its potential for real-time deployment in complex industrial environments. The source code is available at: https://github.com/XiaopengSha/MDEEN.