Facial Wrinkle Segmentation for Cosmetic Dermatology: Pretraining with Texture Map-Based Weak Supervision

Junho Moon,Haejun Chung,Ikbeom Jang

2024-09-13

Abstract:Facial wrinkle detection plays a crucial role in cosmetic dermatology. Precise manual segmentation of facial wrinkles is challenging and time-consuming, with inherent subjectivity leading to inconsistent results among graders. To address this issue, we propose two solutions. First, we build and release the first public facial wrinkle dataset, 'FFHQ-Wrinkle', an extension of the NVIDIA FFHQ dataset. It includes 1,000 images with human labels and 50,000 images with automatically generated weak labels. This dataset could serve as a foundation for the research community to develop advanced wrinkle detection algorithms. Second, we introduce a simple training strategy utilizing texture maps, applicable to various segmentation models, to detect wrinkles across the face. Our two-stage training strategy first pretrain models on a large dataset with weak labels (N=50k), or masked texture maps generated through computer vision techniques, without human intervention. We then finetune the models using human-labeled data (N=1k), which consists of manually labeled wrinkle masks. The network takes as input a combination of RGB and masked texture map of the image, comprising four channels, in finetuning. We effectively combine labels from multiple annotators to minimize subjectivity in manual labeling. Our strategies demonstrate improved segmentation performance in facial wrinkle segmentation both quantitatively and visually compared to existing pretraining methods. The dataset is available at <a class="link-external link-https" href="https://github.com/labhai/ffhq-wrinkle-dataset" rel="external noopener nofollow">this https URL</a>.

Computer Vision and Pattern Recognition,Artificial Intelligence,Machine Learning

What problem does this paper attempt to address?

The paper attempts to address the challenge of accurately segmenting facial wrinkles in cosmetic dermatology. Specifically, manual segmentation of facial wrinkles is both time-consuming and subjective, leading to inconsistent results among different evaluators. To tackle these challenges, the authors propose two solutions:

1. **Constructing and publicly releasing the first facial wrinkle dataset** (FFHQ-Wrinkle): This dataset extends NVIDIA's FFHQ dataset, containing 1,000 images with manual labels and 50,000 images with automatically generated weak labels. This dataset can serve as a foundation for the research community to develop advanced wrinkle detection algorithms.

2. **Proposing a pre-training strategy based on texture maps**: This strategy is applicable to various segmentation models and detects facial wrinkles through a two-stage training method. The first stage is weakly supervised pre-training, using a large amount of weakly labeled data (N=50,000) or mask texture maps generated by computer vision techniques for pre-training. The second stage is supervised fine-tuning, using a small amount of manually labeled data (N=1,000) for fine-tuning. During fine-tuning, the network input includes a combination of RGB images and mask texture maps (four channels).

Through these two methods, the authors aim to reduce the time and cost of manual wrinkle annotation and improve the performance of facial wrinkle segmentation. Experimental results show that this strategy outperforms existing pre-training methods in both quantitative and qualitative aspects.