Research Article
BeautyNet: Joint Multiscale CNN and Transfer Learning Method for Unconstrained Facial Beauty Prediction
Table 1
The proposed BeautyNet architecture dimensions.
| | Name | Filter size/stride, pad | Output size | No. of parameters |
| | Input | — | 120 × 120 × 3 | — | | Conv1 | 5 × 5/1, 2 | 120 × 120 × 96 | 7296 | | MFM1 | — | 120 × 120 × 48 | — | | Pool1 | 2 × 2/2 | 60 × 60 × 48 | — | | Conv2 | 1 × 1/1 | 60 × 60 × 96 | 4704 | | MFM2 | — | 60 × 60 × 48 | — | | Conv3 | 3 × 3/1, 1 | 60 × 60 × 192 | 83136 | | MFM3 | — | 60 × 60 × 96 | — | | Pool3 | 2 × 2/2 | 30 × 30 × 96 | — | | Conv4 | 1 × 1/1 | 30 × 30 × 192 | 18624 | | MFM4 | — | 30 × 30 × 96 | — | | Conv5 | 3 × 3/1, 1 | 30 × 30 × 384 | 332160 | | MFM5 | — | 30 × 30 × 192 | — | | Pool5 | 2 × 2/2 | 15 × 15 × 192 | — | | Conv6 | 1 × 1/1 | 15 × 15 × 384 | 74112 | | MFM6 | — | 15 × 15 × 192 | — | | Conv7 | 3 × 3/1, 1 | 15 × 15 × 256 | 442624 | | MFM7 | — | 15 × 15 × 128 | — | | Conv8 | 1 × 1/1 | 15 × 15 × 256 | 33024 | | MFM8 | — | 15 × 15 × 128 | — | | Conv9 | 3 × 3/1, 1 | 15 × 15 × 256 | 295168 | | MFM9 | — | 15 × 15 × 128 | — | | Pool9 | 2 × 2/2 | 8 × 8 × 128 | — | | Conv10 | 1 × 1/1, 0 | 8 × 8 × 512 | 66048 | | MFM10 | — | 8 × 8 × 256 | — | | Conv11 | 1 × 1/1 | 8 × 8 × 2048 | 526336 | | MFM11 | — | 8 × 8 × 1024 | — | | Res1 | — | 8 × 8 × 1152 | — | | Pool12 | 2 × 2/2 | 4 × 4 × 256 | — | | Pool13 | 2 × 2/2 | 4 × 4 × 128 | — | | Pool14 | 2 × 2/2 | 4 × 4 × 128 | — | | Fc1 | — | 1 × 1 × 512 | 786944 | | MFM12 | — | 1 × 1 × 256 | — | | Drop1 | — | 1 × 1 × 256 | — | | Fc2 | — | 1 × 1 × 5 | 1280 | | Total | | | 2671456 |
|
|
