Automated acne assessment on selfie

• Main Author: Li, Yihang
• Other Authors: Alex Chichung Kot
• Format: Thesis-Master by Research
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Computer science and engineering
• Online Access: https://hdl.handle.net/10356/170532
• Institution: Nanyang Technological University

Deep learning methods are gaining popularity in medical image analysis, while little attention is paid to skin analysis through daily selfie. Although people are getting accustomed to taking care of their beauty and skin quality through daily skincare product rather than dermatology clinics, there are still very few analysis tools for people to compare and evaluate their facial skin quality. Development of deep learning based on facial analysis methods suffers from lack of data and existing approaches are sensitive to pose, light, ethnicity and occlusion, which causes large domain gap. To solve this problem, we proposed a central difference double channel network. In this model, central difference convolutional neural network is deployed to capture gradient information of acne towards common skin. Label distribution learning and multi-task learning skills are also implemented to enhance the learning ability. Meanwhile, we collected a 50-image wild Caucasian acne selfie dataset CauAcne with variable lightness and pose condition. To cope with this challenging dataset, we propose an adapted meta learning network. This meta learning model consists of a 4-way 1-shot data sampler, a 12-layer embedding neural network and a nearest neighbour classifier, in order to make full use of the limited dataset. Transfer learning skills are also used to improve the performance of the adapted meta learning network on CauAcne. After training, the assessment accuracy reached 64% and outperformed the state of the art methods by 30%.