A Comparison of Applying Image Processing and Deep Learning in Acne Region Extraction

Quantifying acne on face images is considered as a challenging task due to the complex skin surfaces, irregular edges and diverse appearances of acnes. A key in this campaign relies upon segmenting lesion areas precisely in the captured images against varying Imaging situation, e.g., illumination, skin condition, imaging angles and etc. To processing these acne data, either theory-driven image processing methods or data-driven deep learning (DL) based methods are commonly utilized in practice. In order to investigate the advantage and shortcoming of the abovementioned two technology roadmaps in quantifying acne task, we develop an image processing pipeline, and make comparison with the state-of-the-art DL methods such as SegFormer, UNETR, Swin-UNet and TransUNet using small data set. The quantitative comparison results reveal that TransUNet performs better in terms of precision, recall, F1-Score and accuracy, whilst image processing methods still have potential in practice due to its annotation-free.