Skin Segmentation Based on Improved DeepLabV3+
Taiyu Han · 2022
In recent years, the number of elderly people and people with disabilities has been increasing.However, the intelligence of the bathing equipment used to assist the elderly and disabled lags far behind that of other medical equipment.The first task of developing an intelligent autonomous bathing aid system without human intervention is to achieve automatic segmentation and localization of skin.The traditional image segmentation model has low accuracy in the case of insufficient light, foam, and water mist occlusion, while the deep learning model exemplified by DeepLabV3+ has high accuracy and robustness.Replacing the backbone feature extraction network in DeepLabV3+ with the lightweight neural network MobileNetV2 can effectively solve the problem of computing power limitation when the deep learning algorithm is deployed to embedded devices, and the generalization performance ofthe improved model is optimized by five scenario-specific data enhancement methods and migration learning ideas.Experiments show that the proposed improved method reduces the weight file size by 93% and 87% with only 1.1% and 1.2% reduction in MIoU compared with the original and Unet models, respectively, achieving a better balance in accuracy and performance.