Defect Recognition of Metal Components Based On Transfer Learning and Feature Fusion

While Artificial Intelligence continues to permeate industrial production, the recognition of surface defects in metal components is constrained by the difficulty of identifying subtle and easily confused surface defects and the lack of a sufficient number of samples for training the recognition model. This study combines five typical neural networks—AlexNet, GoogLeNet, ResNet, Xception, and ResNeXt — by using transfer learning based on the ImageNet dataset, and applies multi-scale convolution, shortcut connection, and mixed pooling to enhance recognition performance. This paper is an algorithm for defect recognition based on transfer learning, feature fusion, and enhanced generalization. The proposed method can accurately recognize defects on the surface of metal components, with a higher accuracy of 97.81% and a true positive rate of 98.32%. Its accuracy on the test dataset varied from 87.69% to 97.81%, which indicates that our model is significantly better than traditional machine learning methods and the original neural network. It can also provide a reference for intelligent defect recognition in metal components.