A Novel Deep Neural Network for Facial Beauty Improvement

Abstract

This study delves into how to combine deep learning and fuzzy logic reasoning to evaluate facial aesthetics and provide targeted makeup recommendations. To further optimize the prediction results, we adopted the BLS method to correct the prediction residuals generated by ResNet-50. Specifically, the predicted appearance score can be expressed as score = p + δ, where p is the predicted result and δ represents the predicted residual of the system. After determining the beauty rating, we further studied four different makeup combinations (x1, x2, x3, x4). Moreover, we introduced fuzzy logic reasoning, defined fuzzy sets and fuzzy relationships, and established membership matrices for each makeup combination. The results of these fuzzy logical reasoning allow us to set a value range of m, n for each makeup method. Based on these reasoning results, we have come up with makeup recommendations for different facial aesthetics. Performance our system with the data collected from internet (accuracy of the calculation = 93.26%), from one volunteer (accuracy of the calculation = 98.14%) and from the both with different makeup skills (accuracy of the calculation = 95.63%) demonstrated that the visual sensing problem is feasible and will be a novel direction for the related engineering applications.