Transfer Learning based Gender Identification using Arbitrary Celebrity Image Sets

Gender Identification is important for security, personalization, and social media analysis, where accurate gender identification enhances the system performance. In this study, we investigated the use of transfer learning for Gender identification from the perspective of minimizing accuracy and efficiency loss. The models, pre-trained on ImageNet weights, were fine-tuned on 300 celebrity images to analyze their performance with varied data constraints. The models were evaluated based on the model architecture and hyperparameters, such as batch size and data split. VGG16 and VGG19 worked impressively with a combined performance level of 98% (97% for female samples and 99% for male samples). However, ResNet50 and ResNet101 showed fluctuating levels of performance, attaining the best accuracy levels of 77.5% and 79.5%, respectively. The results indicate that less complex models, such as VGG16 and VGG19, outperform their more complex versions, such as ResNet50 and ResNet101, on smaller datasets because of their higher efficiency and suitability. More complex models require more sophisticated fine-tuning procedures but tend to have lower performance levels than less complex models. The research identifies that transfer learning significantly eliminates the necessity for longer retraining and customization of models, especially when adapting them to similar tasks. Additionally, considerable performance differences between the male and female categories were identified, highlighting the necessity of balanced datasets and model training to accurately reflect varied gender expressions.