Design and implementation of industrial defect detection system based on EfficientNets

This paper proposes an EfficientNet-based industrial surface defect detection system for automated visual inspection in quality control. The modular design includes three core components: data acquisition/preprocessing, feature extraction, and defect detection/classification. Using EfficientNet-B3 as the backbone, we introduce a multi-scale feature fusion module with adaptive attention to enhance detection of diverse defect sizes and types. Experiments on the NEU dataset and a self-built metal defect dataset demonstrate 95.8% average accuracy, with low computational cost and high inference speed. Ablation studies confirm the fusion module's contribution, boosting accuracy by 3.5%. Deployment on an electronic component production line validates its real-world stability and practicality.