<![CDATA[The suitability of a hairstyle is largely influenced by the shape of a person's face; however, manual identification can often lack consistency due to the observer's subjectivity. This study developed an automated system designed to classify face shapes using the EfficientNet-B4 model, categorizing them into five types: Heart, Oblong, Oval, Round, and Square. The model was trained on a dataset of 27,066 labeled images, which included 19,926 images for training, 3,512 for validation, and 3,628 for testing. The training process involved a two-phase transfer learning approach: first, training the head of the model, followed by fine-tuning the backbone. To enhance performance and mitigate overfitting, data augmentation, learning rate scheduling, and early stopping techniques were utilized. Evaluation results revealed exceptional performance, achieving a validation accuracy of 96. 10%, a test accuracy of 93. 52%, and a macro-F1 score of 0. 935. The highest errors were found in the Oval and Oblong categories, whereas the Square category demonstrated the most consistency. This system is implemented in a web application utilizing Next. js and Express, where face detection is carried out on the client-side using react-webcam and face-api. js. Additionally, the system provides a hairstyle preview to enhance the user experience. Keywords: EfficientNet-B4, face shape classification, hairstyle recommendation, transfer learning, web application.]]>
Copyrights © 2025
