End-to-end face parsing via interlinked convolutional neural networks

Area of Science:

• Computer Vision
• Deep Learning
• Image Segmentation

Background:

• Accurate pixel segmentation of facial parts is crucial for face analysis.
• The two-stage Interlinked Convolutional Neural Networks (iCNN) model is effective but limited by separate training.
• Existing methods require separate training stages, hindering overall performance.

Purpose of the Study:

• To develop a simple, end-to-end face parsing framework.
• To improve the accuracy and performance of the original iCNN model.
• To enable joint training of a two-stage face parsing pipeline.

Main Methods:

• Introduced a Spatial Transformer Network (STN) between the two stages of iCNN.
• Developed the STN-aided iCNN (STN-iCNN) framework for end-to-end joint training.
• Leveraged STN for trainable connections and improved part cropping.

Main Results:

• STN-iCNN significantly improved the accuracy of the original iCNN model.
• Achieved competitive performance on the Helen Dataset for face parsing.
• Demonstrated superior performance and good generalization on the CelebAMask-HQ dataset.

Conclusions:

• The STN-aided iCNN (STN-iCNN) framework enables effective end-to-end joint training for face parsing.
• This approach enhances segmentation accuracy and model generalization.
• The proposed method offers a significant advancement in face parsing technology.