Abstract:Skin aging is an increasing concern, and wrinkle information provides effective feedback on the progress of anti-aging treatments for the skin. It also reflects a person’s lifestyle and offers valuable insights into skin health. Existing wrinkle detection algorithms are influenced by facial features and image backgrounds, require segmentation of the facial region into multiple blocks before detection, and perform well only in the detection of horizontally oriented wrinkles at the forehead, which presents a significant limitation. Motivated by these challenges, this study proposes a facial wrinkle detection algorithm based on an improved DeepLabv3+ model and a semi-automatic labeling strategy, characterized by the following procedures: (1) The algorithm combines facial texture features with rough annotations of wrinkles provided by dermatologists to generate the ground truth required for deep learning; (2) A lightweight, MobileNetV2 network is utilized as the backbone of the model to reduce network parameters and computational complexity. (3) A hybrid attention mechanism is incorporated to enhance feature extraction capability. The deep learning model is subsequently trained using the original images and ground truth labels. The accuracy of the proposed method is evaluated using the Jaccard similarity index. Experimental results demonstrate that the proposed algorithm outperforms traditional algorithms, the U-Net network, the HRNetV2 network, the PAPNet network and the original DeepLabv3+ network in wrinkle detection.