DefinePK

Abstract

Face-based gender recognition system can be used in many applications such as security, human-computer interaction, targeted advertising and social analytics. Finally, we use a new architecture that has higher accuracy but less complexity in this work. First and foremost, we present an EfficientNetB0-MobileNetV2 hybrid model in order to align the strengths of both architectures and achieve a state-of-the-art performance of 97%. For complete compariton, we also tried (1) a pre-trained EfficientNetB2 model with classifier on the top (average accuracy=96.73%); and (2) a CNN-LSTM hybrid, which brings together convolutional feature extraction and sequential processing to yield an average accuracy of 85%. Compared to previous approaches, our proposed EfficientNetB0-MobileNetV2 hybrid architecture takes the advantage of both the classification performance and the model compression, making it better suited for resource-scarce applications in the real world. Through extensive experiments on benchmark datasets, we show that our proposed models not only have a strong robustness, but also a good generalization ability. The research revealed some crucial information about the designing of hybrid deep learning-based gender recognition frameworks which is helping to advance the facial analysis-related researches.