A DEEP LEARNING APPROACH FOR ACCURATE FACIAL WRINKLE SEGMENTATION USING UNET++ MODEL WITH DICE AND FOCAL LOSS FUNCTIONS

Tóm tắt

Facial wrinkle segmentation is essential for various applications in dermatology and cosmetology, yet existing methods often struggle with accurate delineation due to limited dataset diversity and class imbalance. In this paper, we propose a novel facial wrinkle segmentation method based on the Unet++ model, enhanced with dice and focal loss functions. Our approach begins with the construction of an enriched wrinkle dataset sourced from the Flickr-Faces-HQ dataset, ensuring diversity in wrinkle types and complexities. We then introduce a skin region extraction technique to isolate relevant facial areas, enhancing segmentation accuracy. The Unet++ model is employed for wrinkle segmentation, leveraging its encoder-decoder architecture and densely nested skip pathways to capture fine wrinkle details. By integrating dice and focal loss functions, our method effectively addresses class imbalance and improves segmentation performance. Experimental results demonstrate the superiority of our approach in both qualitative and quantitative evaluations, showcasing enhanced wrinkle extraction capabilities and superior segmentation accuracy compared to existing methods. Overall, our study advances the field of facial wrinkle segmentation, offering a robust and reliable method for accurate wrinkle delineation in facial images.