Detection of Human Faces and Prediction of Gender, Age and Emotional States Using Artificial Intelligence Approaches
Master's Thesis
The extraction of diverse personal information from human facial images has gained increasing significance with the rapid advancements in artificial intelligence technologies. In particular, multi-task analysis systems, which enable the simultaneous prediction of multiple attributes from a single facial image, are now widely employed in fields such as human-computer interaction, security, and user profiling. In this study, a comprehensive multi-task prediction system has been developed to estimate information such as emotional state, gender, and age from facial images. In the initial stage of the system, the Haar Cascade method from the OpenCV library was utilized for face detection. For emotion recognition, an EfficientNetV2-based model was designed. To better capture subtle details in facial expressions, the model was enhanced with two different attention mechanisms. Trained on an extended version of the FER-2013 dataset supplemented with additional training samples, the model achieved an accuracy of 82.56%. For gender and age estimation, a preprocessed subset of the IMDb-WIKI dataset’s WIKI partition was employed. In the gender prediction task, a ResNet-50 architecture augmented with an attention mechanism was adopted, yielding successful results with an accuracy of 95.53%. For age prediction, a ConvNeXt V2 based regression model, reinforced with an attention mechanism and equipped with a Spatial Transformer Network (STN) at its input, was utilized, achieving a mean absolute error (MAE) of 4.63. All these tasks were integrated into a real-time operational system with a Haar Cascade-based face detection framework. The results demonstrate the potential of the proposed method to perform versatile and reliable analyses from a single facial image, indicating its applicability in a wide range of domains such as human-computer interaction, security systems, and user profiling.