Table 1 Comparative analysis of image type correlation and size correlation of data set in transfer learning method.
From: A small fishing vessel recognition method using transfer learning based on laser sensors
| Â | Factor | Factor size | Input size | |
|---|---|---|---|---|
Similarity | Size | |||
1 | Similar | Similar | Only one classifier layer is trained using pre-trained weights | When the new data set is small, the fine-tuning of pre-training weights will lead to over-fitting. In the new problem, the image is similar to the original data set, and the features are still related through pre-trained weights |
2 | Similar | Larger | Overall network fine tuning | When the new data set is large enough, there will be no fitting problem in retraining |
3 | Dissimilar | Similar | A few layers can be fine-tuned | Because of the difference between the new data set and the original data set, few convolutional layers need to be retrained to learn features |
4 | Dissimilar | Larger | Transfer learning can fine-tune the whole network | The new data set is large enough and the image difference is large, so transfer learning can be utilized by fine-tuning the whole network |
