Facial beauty prediction via deep learning methods based on SCUT-FBP5500 dataset described in the paper [1] and [2]，which listed at the end of this article, have been partially implemented in this project.
The SCUT- FBP5500 dataset has totally 5500 frontal faces with diverse properties (male/female, Asian/Caucasian, ages) and diverse labels (face landmarks, beauty scores within [1, 5], beauty score distribution), which allows different computational models with different FBP paradigms.
Further more, three recently proposed CNN models with different structures for FBP, including AlexNet, ResNet-18 and ResNeXt-50, which are trained by initializing weights using networks pre-trained on the ImageNet dataset, have been evaluated on the dataset.
The results illustrates that the deepest CNN-based ResNeXt-50 model obtains the best performance. The Pearson correlation coefficient is 0.8997.
The second paper, recasts facial attractiveness computation as a label distribution learning problem and puts forward an end-to-end attractiveness learning framework. Extensive experiments are conducted on a standard benchmark, the SCUT-FBP dataset, where shows significant advantages over other state-of-the-art work.
I have trained Label distribution learning model based on ResNet fine-tuing, you could refer the paper to get detail infomation. The Pearson correlation coefficient is 0.95, which is tested on the SCUT-FBP5500 dataset.
- Python 3.x
- Tensorflow
- Keras
- numpy
- opencv
- h5py
- dlib
- The trained model files could be downloaded via follow links:
- put Label distribution learning model under inference/ldl+resnet folder, and run beauty_predict.py
If you want to train your own model, read the follow part.
- Prepare Data
- download dataset, unzip and put it under Project Root folder
- cd train/ldl+resnet and run prepare_data.py to prepare data before training label distribution learning model
- Train and test
- run train_model.py under train/ldl+resnet folder, this maybe take a long while, which depends on your machine. The model file named model-ldl-resnet.h5 will be generated in the same folder.
- run test_model.py script, which will test the model on test dataset, and print the label score and predict score for each image in the test dataset, finally the The Pearson correlation coefficient has been listed in the last line.
- Predict
- copy the model file nameed model-ldl-resnet.h5 to inference/ldl+resnet folder, you could call beauty_predict.py to use the model now.