| Summary: | Gender classification is one of the key features in soft biometrics besides age, ethnicity, facial expression, etc. Gender classification based on skin tone has its own importance that can further improve the performance of facial recognition systems. Most Convolutional Neural Network (CNN) models require a large amount of training data to improve classification accuracy and increase processing time. Fortunately, the MobileNetV2 model overcomes this problem by running faster than the other models. However, the model’s accuracy suffers when the gender classification results based on skin tone reach 50% accuracy, indicating that the model suffers from an “overfitting” problem. To address this issue, the proposed research constructed a novel face images dataset containing 6250 face images that chosen from original FaceARG dataset and divided equally into Bright (3125) and Dark (3125) skin tone. Each skin tone (Bright and Dark) is equally divided into two genders, 1563 (male) and 1563 (female). The new FaceARG dataset is then used to run two types of experiments (Bright and Dark) on the MobileNetV2 model. The Fine-Tuning method from Transfer Learning is then applied to the MobileNetV2 model, along with method gains from previous studies. The Dark experiment achieved the highest accuracy on training dataset, which is 97.4%, compared to 50% on the model without fine-tuning, and the Bright experiment achieved the highest accuracy on test dataset, which is 89.8%, compared to 50% on the model without fine-tuning. The findings of this study will determine the ability of the proposed model to accurately classify gender based on skin tone. © The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2024.
|
|---|