An evaluation method for model transfer learning performance in industrial surface defect detection tasks

Transfer learning has become one of the most effective techniques to reduce the supervision cost of learning tasks, and has been applied in various domains. However, how to accurately and efficiently transfer knowledge between different domains is a challenging issue. Some previous pilot work can evaluate the transfer performance of different domains, but the practical performance of its application cannot be guaranteed when it is applied to the engineering domain. To tackle this issue, we focused on the task of surface defect detection in industrial engineering. The proposed method Defect-deepscore (D-deepscore) could quickly and accurately evaluate models from different source domains on a target domain, and then select the source domain model without any fine-tuning process. D-deepscore takes the parameters from deeper layers in the convolutional neural network, which are further processed by dimensionality reduction and information correlation analysis. In the experiments we demonstrate that finetuning the commonly used ImageNet pre-trained model is not necessarily the best choice and transfer learning from defect dataset will be more effective. Then, we evaluate the multiple pretrained models which were trained on multiple surface defect datasets, the results show that there is a strong correlation between D-deepscore’s model evaluation scores and the classification accuracy. By comparing with existing SOTA (State-of-the-art) methods that focus on model transfer learning performance, D-deepscore improves the evaluation accuracy by 49.9 % over the best previous work. The proposed D-deepscore could provide a fast selection of the best pre-trained model for industrial defect detection tasks, which ultimately leads to improved detection performance.