Deep Transfer Learning For Abnormality Detection
Jie Kong, Yonghui Xu, Han Yu · 2019
Deep learning has proven to be effective in learning scenarios with massive training data. However, in many real applications (i.e., abnormality detection), there is a lack of sufficient data to achieve a good deep learning model. Considering the fact that collecting massive labeled training data for a new task is often expensive and time-consuming, it is critical to transfer and reuse the knowledge of the labeled data or simulation data to the new task. To tackle this issue, we propose a deep transfer learning method in this paper. On one hand, we pre-train a basic deep learning model with the related training data. Then, we treat the learning model as a starting point for the current problem to train the new deep learning model. On the other hand, we utilize the Generative Adversarial Nets (GAN) in the learning process to transfer knowledge from simulation data and further enhances the discriminative power of the model. Besides, we apply the proposed method to abnormality detection problem. Experiments in Bone X-Ray anomaly detection show that the proposed deep transfer model can significantly improve performance compared to the basic deep learning model.