Real-time Industrial Surface Defect Detection Based on Lightweight Convolutional Neural Networks

Abstract

Industrial surface defect detection represents a critical component in manufacturing quality control systems, demanding both high accuracy and real-time performance. Traditional computer vision approaches often struggle with computational complexity and inference speed requirements in production environments. This paper presents a novel lightweight convolutional neural network architecture specifically designed for real-time industrial surface defect detection applications. The proposed method integrates advanced model compression techniques, multi-scale feature extraction modules, and attention mechanisms to achieve optimal balance between detection accuracy and computational efficiency. Experimental validation on multiple industrial datasets demonstrates superior performance compared to existing approaches, achieving 94.7% detection accuracy with inference times of 12.3ms on edge computing devices. The developed framework addresses key industrial requirements including robustness to lighting variations, multi-class defect recognition, and deployment feasibility in resource-constrained environments. Implementation results across various manufacturing scenarios validate the practical applicability and scalability of the proposed solution for real-world industrial deployment.