((new)): Deep Cheeks 2

Facial analysis has progressed dramatically with deep convolutional neural networks (CNNs), yet most works treat the face as a monolithic entity. The —covering the area from the zygomatic arch to the mandibular angle—is of particular clinical and commercial interest. Dermatologists assess cheek skin for erythema, hyperpigmentation, and lesion distribution, while cosmetic applications rely on cheek plumpness and skin‑tone uniformity for product recommendation.

Extensive experiments demonstrate that Deep Cheeks 2 outperforms state‑of‑the‑art facial segmentation models (U‑Net, DeepLab‑v3+, BiSeNet) by on the cheek region and +12.4 % Pearson correlation on aesthetic scores. Ablation studies confirm the importance of each architectural component. We release the source code, trained weights, and the CheekWILD‑2 dataset under an open‑access license. Deep Cheeks 2