1 Introduction
2 Related Work
3 Preliminary
3.1 Maximum Mean Discrepancy
3.2 Reproducing kernel Hilbert space
4 Transfer Learning for Image Emotion Analysis
4.1 Joint Maximum Mean Discrepancy
4.2 Deep Transfer Learning Model
5 Experiment
Dataset | Amusement | Anger | Awe | Contentment | Disgust | Excitement | Fear | Sadness | Sum |
---|---|---|---|---|---|---|---|---|---|
FI | 4861 | 1236 | 3055 | 5292 | 1616 | 2827 | 998 | 2815 | 22,700 |
ArtPhoto | 101 | 77 | 102 | 70 | 70 | 105 | 115 | 166 | 806 |
IAPS-Subset | 37 | 8 | 54 | 53 | 74 | 55 | 42 | 62 | 395 |
- CTD [29]: The CNN model is fine-tuned only with labeled data in target domain. This is the basic method used for image emotion classification.
- CBD: The model is fine-tuned with labeled data in both the source and the target domain without transferring modules.
- DAN [14]: This is a classical deep transfer learning method, which measure the domain distribution discrepancy with MMD and reduce it in a layer-wised way.
- JAN [16]: This method aligns full-conncted layers of a CNN and minimize their joint distribution discrepancy with JMMD.
CTD | CBD | DAN | JAN | Ours | |
---|---|---|---|---|---|
\(\mathbf{F }\rightarrow \mathbf{I }\) | 24.81 | 26.30 | 25.93 | 27.78 | 29.63 |
\(\mathbf{A } \rightarrow \mathbf{I }\) | 22.96 | 24.44 | 30.00 | 27.41 | 27.04 |
\(\mathbf{F }\rightarrow \mathbf{A }\) | 39.66 | 36.92 | 36.24 | 36.75 | 37.61 |
\(\mathbf{I } \rightarrow \mathbf{A }\) | 30.77 | 34.56 | 36.78 | 34.81 | 39.15 |