Research Article
BeautyNet: Joint Multiscale CNN and Transfer Learning Method for Unconstrained Facial Beauty Prediction
Table 7
Performance comparison of state-of-the-art methods.
| Method | Transfer learning | Classification accuracy (%) | Pearson’s correlation coefficient |
| Raw pixel [15] | No | 48.11 | — | Eigenfaces [15] | No | 46.52 | — | CRBM [15] | No | 51.62 | — | K-means [15] | No | 52.54 | — | Multiscale K-means [15] | No | 55.07 | — | NIN_Imagenet [45] | No | 55.60 | 74.69 | NIN_Imagenet [45] | Yes | 58.30 | 76.96 | DeepID2 [46] | No | 55.90 | 73.75 | DeepID2 [46] | Yes | 60.25 | 76.57 | GoogLeNet [47] | No | 57.20 | 78.14 | GoogLeNet [47] | Yes | 62.59 | 79.78 | VGG_CNN_S [48] | No | 57.30 | 78.87 | VGG_CNN_S [48] | Yes | 62.69 | 81.12 | LightenNet | No | 64.36 | 79.41 | LightenNet | Yes | 65.82 | 81.59 | BeautyNet | No | 64.84 | 80.20 | BeautyNet | Yes | 67.48 | 83.54 |
|
|