Penerapan Swin Transformer Pada Klasifikasi Bentuk Wajah Untuk Rekomendasi Bentuk Kacamata

Abstract

Facial shape serves as an identity marker that reflects the unique physical characteristics of each individual. In general, facial shapes are categorized into five main types: square, round, oval, heart, and oblong. Information regarding facial shapes can be utilized in various aspects of personalization, including the selection of suitable eyeglass models. As the fashion industry evolves, eyeglasses are no longer solely functional as vision aids but also serve as aesthetic elements that enhance one’s appearance. Therefore, selecting eyeglass models based on facial shape is a crucial factor in improving an individual's visual appeal. In online shopping, customers often face difficulties in determining the most suitable eyeglass model. Thus, an AI-based automated system is required to provide optimal recommendations. This study implements the Swin Transformer model to classify facial shapes as a basis for eyeglass model recommendations. Swin Transformer is a Deep Learning architecture that employs a hierarchical patch-based image partitioning approach and processes images through a series of transformer blocks in a progressive manner. The results of this study demonstrate that the model performs well in classification tasks. With a learning rate of 0.0001, a batch size of 32, and 64 epochs, the model achieved a training accuracy of 99.15%, a validation accuracy of 98.47%, and a test accuracy of 97.63%. The small accuracy difference (0.68%) indicates that the model does not exhibit significant overfitting. Furthermore, the consistently low loss values across all data subsets indicate the model’s stability. These findings confirm that the developed system is capable of accurately classifying facial shapes and providing appropriate eyeglass recommendations.