Skin blemishes generation on facial images based on computer vision

• Main Author: Wu, Yiran
• Other Authors: Alex Chichung Kot
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Engineering::Computer science and engineering
• Online Access: https://hdl.handle.net/10356/173395
• Institution: Nanyang Technological University

People often suffer from facial skin blemishes, such as acne and pigmentation, but most do not have the pathological knowledge to distinguish the symptoms, so automated identification and classification of spots on the skin can assist in skin care and analysis. Deep Learning and Computer Vision have been widely used for image semantic segmentation, and the models have become mature, which can be used to automate the recognition of skin spots. Segmentation models need to be trained on relevant datasets, in this task the images of spots on the facial skin are needed. However, no professionally labeled open-source datasets of blemishes on the facial skin can be found, so this project proposes to increase the data through data synthesis, which is used to improve the segmentation effect. Our industry partner provided full-face images that have professionally labeled blemishes. The author applied the DermGAN [1] model and Palette [2] model to the inpainting method, and based on the provided images, the author made local spot datasets applicable to these models to carry out spot generation. A complete process is given to generate a single spot in any part of the full-face image. The author successfully produced datasets suitable for this project and trained the model to generate a local spot image with high credibility in visual effects. The author used the inpainting method to improve the fit between generated local images and the background full-face image and made progress in visual effects, but there is still much room for improvement in perfectly integrating local images into the background.