Evaluating the Detection Performance of Different Convolutional Neural Network Models for Steel Surface Defects

To ensure quality control in the steel industry, accurate and efficient surface defect detection technology is essential. Deep learning, especially Convolutional Neural Networks (CNNs), has shown great potential in this field, but there is still a lack of systematic performance comparison of models with different architectural concepts under a unified framework. This study comprehensively evaluates four representative CNN models - VGG16, ResNet50, InceptionV3, and MobileNetV2 - based on the steel surface defect dataset publicly available from Northeastern University (NEU-CLS). This study adopts a unified data preprocessing, data augmentation strategy, and a fine-tunable unified transfer learning method when conducting experiments. The performance of the models is then evaluated by comparing the accuracy, precision, recall, F1 score, and confusion matrix of each model. The experimental results show that all models have very good performance in the task of steel surface defect classification, with test accuracy exceeding 97%. Among all models, VGG16 has the highest accuracy of 99.26%. The lightweight MobileNetV2 has the same accuracy as the more complex ResNet50, both reaching 98.52%. This result shows that MobileNetV2 has achieved a good balance between performance and computational efficiency. In short, the main contribution of this study is to provide a judgment benchmark for future model selection. When weighing the model complexity and classification accuracy, you can make a choice based on the results of this experiment.