Sobel-Enhanced Lightweight CNN for Real-Time Industrial Defect Detection on Embedded Edge Platforms

Abstract

CNN-based automated defect detection is frequently used in industrial environments, but direct training on raw RGB images can degrade model performance in low-resolution or visually challenging conditions. Therefore, lightweight and robust preprocessing conditions are necessary for model training. To address this problem, a lightweight CNN model is proposed, and this model is further improved with Sobel optimization using 3 × 3 and 5 × 5 cores to facilitate edge detection on RGB images. The results show that the gradient information obtained from Sobel improves classification performance compared to the basic RGB model. In particular, the 5 × 5 Sobel configuration achieves the highest test accuracy and can capture more missed defect types for the defective class. Furthermore, the confusion matrix of the 5 × 5 Sobel configuration model confirms better class separation for subtle defects. To evaluate the models' performance under real-world conditions, they were integrated into a Raspberry Pi using ONNX. In real-time tests, the models demonstrated operation above 20 FPS. In general, Sobel, a gradient-based edge preprocessing algorithm, improves CNN defect detection performance without increasing model complexity, demonstrating its usability in resource-constrained applications. © 2026 IEEE.