Abstract
<title>Abstract</title> <p>This study utilizes multi-light facial imaging to develop a lightweight Convolutional Neural Network (CNN) architecture based on MobileNetV2. The dataset consists of 211 participants, each with eight images captured under the following light modalities: white light, blue light, red light, brown light, UV light, Wu light, positive polarization, and negative polarization. Each respondent also has 13 parameters that quantitatively represent skin conditions, including sebum, pores, spots, wrinkles, acne, blackheads, dark circles, skin color, PL sensitivity, UV spots, pigment, UV acne, and collagen fibers. The eight image datasets were processed through resizing, pixel normalization, data augmentation, and target normalization. The MobileNetV2 model was designed to extract light modalities via feature fusion and perform quantitative predictions of skin parameters. Model evaluation was conducted using 5-fold cross-validation, parameter distribution analysis, prediction-actual correlation, regression, and Bland-Altman plots. The multi-light method approach has the potential to provide more visual information than standard RGB images. The performance of the sebum and dark circle parameters is relatively lower than that of other parameters. The uneven distribution of the dataset limits the performance of the lightweight CNN model. The low R² value indicates that the development of prediction models based on lightweight CNNs still faces challenges in feature extraction. Thus, this study provides an initial baseline for the development of skin analysis systems using a multi-light approach. This study also highlights the need for a well-balanced dataset with a large sample size in future research.</p>