Multi-Model Convolutional Neural Network Fusion for Enhanced Metal Surface Defect Detection and Classification

Abstract

The rapid growth of China's non-ferrous metal industry highlights the increasing demand for high-precision defect detection in metal structures used in aerospace, automotive, and high-speed rail industries. Traditional defect detection methods, such as eddy current, magnetic leakage, infrared, and ultrasonic techniques, often face limitations in material adaptability and defect classification accuracy. In response to these challenges, this study proposes a novel approach leveraging multi-model convolutional neural network (CNN) fusion for metal surface defect detection. By optimizing existing CNN architectures and integrating a feature fusion algorithm, the proposed method effectively classifies 12 types of metal surface defects, including scratches, orange peel, and bottom leakage. Transfer learning is employed to train single CNN models, which are subsequently fused to enhance classification accuracy. Experimental results demonstrate that the multi-model fusion network outperforms individual CNN models in terms of accuracy, recall rate, and F1 score, validating the superiority of this approach for comprehensive metal defect detection.