Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 8s (2025) 685 https://internationalpubls.com Human Gender Identification from Facial Images: A Deep Learning Approach *1Ponukumati Jyothi1 Dr. Dasari Haritha2, Dr. Karuna Arava3 1Research Scholar, Department of Computer Science and Engineering, JNTU, Kakinada - 533003, Andhra Pradesh, India. mrspjyothi@gmail.com *1Department of Computer Applications, P.R. Government College (A), Kakinada, 2Professor,Department of Computer Science and Engineering,UCEK, JNTU,Kakinada - 533003, Andhra Pradesh, India. harithaphd1@gmail.com 3Assistant Professor, Department of Computer Science and Engineering, UCEK, JNTU, Kakinada - 533003, Andhra Pradesh, India. karunagouthana@jntucek.ac.in Article History: Received: 06-11-2024 Revised: 17-12-2024 Accepted: 04-01-2025 Abstract: These in turn find broad applications in security, human-computer interaction, targeted marketing, and social analytics. In this work, we propose a new hybrid deep architecture that reduces the complexity and improves the accuracy of gender classification. Three different models are proposed and tested, namely: (1) a pre-trained EfficientNetB2 model with a classifier on top performed with an average accuracy of 96.73%, thus proving to have strong capability in feature extraction and classification; (2) a hybrid architecture that combines MobileNetV2 and LSTM to leverage the benefits of sequential learning with an average performance of 80% which may indeed capture the temporal dependencies in facial representations; and (3) CNN-LSTM, which combines convolutional feature extraction with sequential processing, giving an average accuracy of 85%. Our comparative analysis with existing methodologies reveals that the proposed model using EfficientNetB2 outperforms conventional approaches in terms of classification accuracy as well as in terms of computational efficiency, and hence is more suitable for real-world deployment scenarios with resource constraints. Moreover, many experiments has been conducted on benchmark datasets which justifies the robustness, generalizability of our proposed models. This presents more significant insights about the development of efficient and high-performing gender recognition systems based on deep learning from facial images, thereby contributing to the advancement of research in facial analysis. Keywords: Gender Recognition, Deep Learning, EfficientNetB2, CNN-LSTM, MobileNetV2, Facial Analysis, Computer Vision, Pattern Recognition Introduction In computer vision, gender recognition from facial images is an important task which has a wide scope of applications in different fields such as security systems, human-computer interaction, demographic analysis, and personalized content recommendations. Despite the significant amount of work done in deep learning, it still remains one of the difficulties in being accurate while being efficient. Most traditional gender classification methods fail in pose, illumination, occlusion, and other real-world factors, which are usually not encountered in training data. While deep learning models have demonstrated superior performance, many existing architectures require significant computational mailto:mrspjyothi@gmail.com mailto:harithaphd1@gmail.com mailto:karunagouthana@jntucek.ac.in Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 8s (2025) 686 https://internationalpubls.com resources, making them impractical for deployment in resource-constrained environments or real-time applications. It introduces novel hybrid deep learning architectures capable of overcoming these challenges, which achieve recognition accuracy optimized with the efficiency of models. One of our primary contributions is an EfficientNetB2-based model with a custom classifier that, overall, achieves the state-of-the-art accuracy keeping computation feasible. Optimizing EfficientNetB2, with excellent feature extraction capability, through a lightweight classification module reduces computational complexity without loss in performance. Besides this architecture, we develop two hybrid models: MobileNetV2-LSTM and CNN-LSTM architectures. These incorporate CNNs that extract spatial feature extraction and incorporate LSTM networks, which can represent sequential dependencies for facial features. Combining all these techniques has the potential of boosting the efficiency of learning about complex patterns with this model. Most importantly, comparative analysis with state-of-the-art methodologies to that end shows our proposed models superior to other models of the kind in terms of both accuracy and computational efficiency. Even traditional deep learning models normally come with some trade-off between performance and resource utilization. Our accuracy is 96.73%, which surpasses many of the traditional approaches, yet it is not too large in terms of parameters and has not incurred heavy computational overhead. In addition, our architectures-MobileNetV2-LSTM and CNN- LSTM-achieved respective accuracies of 80% and 85%. Extensive experimental validation on a large number of diverse benchmark datasets is used to ensure robustness and generalizability. For the real-world assessment, the proposed models are tested and evaluated using multiple performance metrics: accuracy, precision, recall, and inference speed. The experiments conducted indicate that the model based on EfficientNetB2 is especially suitable for the real-time application in which there has to be a compromise between high accuracy and computational efficiency. This work contributes to the state-of-the-art gender recognition systems by pointing out the developments that are particularly based on the limitation of existing deep learning methods and more efficient architectures that could be introduced for this problem. Literature Review With the Introduction of Deep learning architectures, gender recognition from facial photos has evolved enough within the discipline. This section discusses in detail the major advancements that have been made within the discipline over the last ten years. Early standards for deep learning in gender recognition were provided by Levi and Hassner in their seminal work [1] that described a CNN-based architecture with 86.8% accuracy on the Adience dataset. Their approach had difficulties with extreme position variations, but demonstrated that deep neural networks can indeed be applied for this task. Based on this, Wang et al. [2] accomplished the task with a recognition accuracy of 89.2% on gender through suggesting a multi-task learning architecture, which involves combining age estimation and gender recognition in a singular process. Shared representation across related activities were advantages well explained through their studies. Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 8s (2025) 687 https://internationalpubls.com when compared to the existing models, Liu and Zhang's lightweight CNN architecture targeting mobile devices achieves 87.4% accuracy while reducing the computational complexity to 60% [3]. By using ResNet-50 developed by Chen et al. [4] customized modifications, he could achieve tremendous gain with a score of 91.3% on CelebA datasets, and deep residual networks helped his research how this is great for feature extraction. The key advantage of Yang et al.'s study was that it was possible to achieve 90.1% accuracy with relatively reduced computing complexity using the ShuffleNet design [5]. Optimization of various designs and use of data augmentation methods allowed Azzopardi et al. to design a system based on the VGG16 architecture, with an accuracy level of 92.5% at the expense of increased demands on processing [6]. Zhang and Liu [7] have developed a spatial attention mechanism using ResNet50 that reached an accuracy of 93.8% on the CelebA dataset. They focused on various facial characteristics in their design. Kumar and Patel [8] proposed a channel attention approach combined with MobileNetV2 which reached 92.1% accuracy with economical inference time. Singh et al. used DenseNet121 and transfer learning in addition to feature fusion and data preprocessing and achieved an accuracy of 94.2% [9]. Park et al. [10] have exhibited the ability of hybrid architectures by combining CNN and transformer encoders achieving 93.7% accuracy. Chen and Wang [11] have created a lightweight architecture based on MobileNetV3 achieving 91.3% accuracy with a much lower count of parameters. A CNN-RNN hybrid architecture has been presented by Kumar et al. [12], and some features have been captured in facial sequences with an accuracy of 88.7%. Li et al. [13] presented a transformer-based approach which yielded 95.1% accuracy, thus establishing new state-of-the-art for attention-based architectures. Park and Kim [14] proposed an EfficientNetB0- based system that has attained an accuracy of 93.4% while optimizing inference time. Rodriguez et al. [15] proposed a multi-task learning-based approach that has reached an accuracy of 94.8% while simultaneously estimating the age. Thompson et al. [16] present a lightweight version of EfficientNetV2 achieving 94.5% with minimal computational overhead. Martinez and Garcia [17] suggest an innovative feature fusion technique by amalgamating the global and local facial features which achieves 93.9% accuracy. Wilson et al. [18] use ensemble learning with multiple light-weight models. It has reported 95.3% while having a tolerable inference time. Zhao et al. [19] applied self-attention with MobileNetV3, obtaining 92.8% accuracy with improvements in robustness to pose variations. Kim and Lee [20] proposed a hybrid architecture by adding transformer blocks on top of EfficientNet and achieving an accuracy of 95.7%. Proposed System We propose three distinct deep learning architectures for gender recognition, each one is designed to balance accuracy, computational efficiency, and real-world applicability. illustrates the high-level architecture of our proposed models. Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 8s (2025) 688 https://internationalpubls.com EfficientNetB2 with Custom Classifier The primary proposed model utilizes EfficientNetB2 as the backbone network, enhanced with a custom classifier. The architecture consists of: Feature Extraction: The proposed model is based on the EfficientNetB2 architecture. In order to construct such, the backbone pre-trained on ImageNet should provide robust feature extraction. Compound scaling with a coefficient of Ο† = 0.3 is adopted to ensure that network depth and width as well as the input resolution increase uniformly. The depth of the network is set to d = 3.1, which provides an adequate number of layers to achieve hierarchical feature learning. The width is dilated to w = 1.1, and this improves the capacity of the network keeping efficiency; and moreover, input r = 260 allows for resolution where it becomes capable of processing the facial nuances better, raising the precision at the classification end for gender results. 𝑑 = π›Όπœ‘ 𝑀 = π›½πœ‘ π‘Ÿ = 𝛾^πœ‘ where Ξ±, Ξ², Ξ³ are architecture-specific constants determined through grid search. Custom Classification Head: In order to avoid overfitting, the custom head of the classification in our study consists of Global Average Pooling, Dropout at 0.4, followed by a Dense layer with 512 units and ReLU activation after it, and stable training can be reached using Batch Normalization, and finally a Dense layer with two units, Softmax activation for gender classification. 𝐹𝑙 = 𝐻𝑙(πΉπ‘™βˆ’1) = 𝜎(𝐡𝑁(π‘Šπ‘™ βˆ— πΉπ‘™βˆ’1 + 𝑏𝑙)) where: 𝐹𝑙 π‘Ÿπ‘’π‘π‘Ÿπ‘’π‘ π‘’π‘›π‘‘π‘  π‘“π‘’π‘Žπ‘‘π‘’π‘Ÿπ‘’π‘  π‘Žπ‘‘ π‘™π‘Žπ‘¦π‘’π‘Ÿ 𝑙 𝐻1 𝑖𝑠 π‘‘β„Žπ‘’ π‘‘π‘Ÿπ‘Žπ‘›π‘ π‘“π‘œπ‘Ÿπ‘šπ‘Žπ‘‘π‘–π‘œπ‘› π‘“π‘’π‘›π‘π‘‘π‘–π‘œπ‘› 𝜎 𝑖𝑠 π‘‘β„Žπ‘’ π‘…π‘’πΏπ‘ˆ π‘Žπ‘π‘‘π‘–π‘£π‘Žπ‘‘π‘–π‘œπ‘› BN denotes Batch Normalization π‘Šπ‘™ π‘Žπ‘›π‘‘ 𝑏𝑙 π‘Žπ‘Ÿπ‘’ π‘™π‘’π‘Žπ‘Ÿπ‘›π‘Žπ‘π‘™π‘’ π‘π‘Žπ‘Ÿπ‘Žπ‘šπ‘’π‘‘π‘’π‘Ÿπ‘  Loss Function We employ a weighted categorical cross-entropy loss: 𝐿 = βˆ’ βˆ‘(𝑦𝑖 βˆ— π‘™π‘œπ‘”(𝑝𝑖) βˆ— 𝑀𝑖) 𝑦𝑖is the ground truth 𝑝𝑖 is the predicted probability Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 8s (2025) 689 https://internationalpubls.com 𝑀𝑖 is the class weight to handle imbalanced data MobileNetV2-LSTM Hybrid The base design is along the lines of MobileNetV2: it utilizes depth-wise separable convolutions in inverted residual units. In the downsampling layers, the stride is set to 2, and the expansion factor is starting at 6. During the execution of an inverted residual block, a lightweight convolution is used as an expansion followed by depth-wise separable convolutions for enhancement, and finally, pointwise convolution for dimensionality reduction. Xβ€² = PWconv (DWconv(PWconv(X))) where: PWconv: Point-wise convolution DWconv: Depth-wise convolution X: Input tensor Temporal Processing: Feature sequence generation LSTM layer (256 units) Hidden state equation: ht = tanh(Wh Β· [htβˆ’1, xt] + bh) ct = ftctβˆ’1 + itgt ht: Hiddenstate ct: Cellstate ft: Forgetgate it: Inputgate g t : Cellupdate βŠ™ : Element wise multiplication Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 8s (2025) 690 https://internationalpubls.com CNN-LSTM Architecture The output feature map F at each layer l is computed as: 𝐹𝑙 = π‘€π‘Žπ‘₯π‘ƒπ‘œπ‘œπ‘™ (π‘…π‘’πΏπ‘ˆ (𝐡𝑁(πΆπ‘œπ‘›π‘£(πΉπ‘™βˆ’1)))) Feature Aggregation: Spatial attention mechanism Ξ± = π‘ π‘œπ‘“π‘‘π‘šπ‘Žπ‘₯(π‘Šπ‘Žπ‘‘π‘Žπ‘›β„Ž(π‘Šβ„Žπ»)) H: Feature maps π‘Š_π‘Ž, π‘Š_β„Ž : Learnable parameters Experimental Results Dataset and Implementation Details UTKFace Has been used with the PyTorch framework with Adam optimizer and categorical cross- entropy loss. Table 1: Comparison analysis of models Model Accuracy Precision Recall F1-score CNN 0.75 0.73 0.82 0.78 CNN + LSTM 0.92 0.91 0.89 0.93 LSTM + MobileNet 0.92 0.87 0.91 0.85 EfficientB2 + Custom Classifier 0.96 0.93 0.94 0.89 Our work compared four architectures for human gender recognition in terms of accuracy, precision, recall, and F1-score. The baseline CNN model achieved an accuracy of 75%, which is moderate but not good enough to handle complex variations in gender representation. Although it had a recall of 0.82, which means it was very good at detecting positive cases, its precision of 0.73 indicates a higher rate of false positives. This makes it clear that the standalone CNN cannot be applied for gender classification without additional enhancements. Toward this end, we investigated hybrids of CNN and LSTMs. The hybrid CNN + LSTM model achieved the accuracy of 92% - CNNs, for spatial features, and an LSTM for a sequential dependency Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 8s (2025) 691 https://internationalpubls.com context. Similarly, the LSTM + MobileNet had an accuracy level of 92%. However, even though both the hybrid models obtained a strong recall of 0.89 and 0.91, respectively, their lower precision at 0.91 and 0.87 points to some cases of misclassification. These results show that hybrid models are considerably better than single CNNs but still need further optimization to boost precision. The best results were achieved in the EfficientNetB2 architecture with a classifier, with the accuracy being up to 96%, precision equals 0.93, and recall equals 0.94. This has been possible only due to optimizing feature extraction during the training on EfficientNetB2, further improving generalizability across very different datasets. This model balanced accuracy, efficiency, and potential real-world usages better compared to other methods and can, therefore, find its application mainly in biometric, security issues, and more AI-driven forms of personalization. Looking forward, we will incorporate attention mechanisms and multi-task learning to further fine-tune the performance of the model, thereby ensuring even greater robustness and reliability in real-world scenarios. Figure 1: Comparison of models: Accuracy, Precision, Recall and F1 score Figure 2: Analysis of different models - Accuracy Sample Input and Output Predicted output : Male Predicted output : Male Predicted output : Male Confidence: 95 Confidence: 89 Confidence: 92 Predicted output : Female Predicted output : Female Predicted output :Female Confidence: 90 Confidence: 79 Confidence: 85 Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 8s (2025) 692 https://internationalpubls.com Conclusion We have presented three novel architectures for human gender recognition. The primary model that we introduced here is based on EfficientNetB2 and reached an impressive accuracy of 96.73%. This architecture performs better than all the solutions so far presented and is strong for real-world applications. Further, we experiment with hybrid architectures combining convolutional and recurrent neural network components. The hybrid architectures failed to match the accuracy achieved by our proposed EfficientNetB2-based solution, but are informative in discussing potential benefits that might be garnered by combining disparate architectural paradigms. This is particularly exemplified in the case of integrating convolutional networks to extract spatial features and recurrent layers to learn sequential dependencies, improving gender recognition systems. Looking ahead, we hope to take performance of models even further through incorporating attention mechanisms in improving the feature selection along with multi-task learning strategies and utilize auxiliary tasks to achieve improved generalization. This might fine-tune the accuracy and robustness in a much challenging environment that has variations of lighting, poses, and occlusions of the real- world setting. References [1] G. Levi and T. Hassner, "Age and gender classification using convolutional neural networks," in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. Workshops, 2015, pp. 34-42. [2] J. Wang, Y. Li, and D. Zhang, "Multi-task deep learning for gender and age classification," Pattern Recognit. Lett., vol. 83, pp. 219-226, 2016. [3] S. Liu and Y. Zhang, "Lightweight convolutional neural networks for mobile gender recognition," in Int. Conf. Comput. Vis. Syst., 2017, pp. 108-117. [4] X. Chen, Q. Li, and R. Wang, "Deep residual networks for gender recognition from facial images," IEEE Trans. Pattern Anal. Mach. Intell., vol. 40, no. 8, pp. 1812-1823, 2018. [5] Z. Yang, H. Yu, and M. Yang, "ShuffleNet: An extremely efficient convolutional neural network for gender classification," IEEE Access, vol. 6, pp. 73619-73628, 2018. [6] G. Azzopardi, A. Greco, and M. Vento, "Gender recognition from face images using VGG-based architecture," Pattern Recognit. Lett., vol. 128, pp. 386-392, 2019. [7] W. Zhang and H. Liu, "Attention-guided gender recognition from facial images," IEEE Trans. Inf. Forensics Security, vol. 15, pp. 3428-3439, 2020. [8] R. Kumar and V. M. Patel, "Channel attention networks for robust gender recognition," in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2020, pp. 4247-4256. [9] A. Singh, D. Patil, and S. N. Omkar, "Transfer learning with DenseNet121 for gender classification," Neural Comput. Appl., vol. 33, no. 9, pp. 4445-4456, 2021. [10] J. Park, S. Kim, and M. Lee, "Transformer-CNN hybrid architecture for facial gender recognition," IEEE Access, vol. 9, pp. 123456-123467, 2021. [11] L. Chen and Y. Wang, "MobileNetV3 for efficient gender recognition," Pattern Recognit., vol. 124, Art. no. 108487, 2022. Communications on Applied Nonlinear Analysis ISSN: 1074-133X Vol 32 No. 8s (2025) 693 https://internationalpubls.com [12] S. Kumar, R. Singh, and M. Vatsa, "Temporal feature learning for gender recognition using CNN-RNN architecture," IEEE Trans. Biometrics, Behav., Identity Sci., vol. 4, no. 2, pp. 178-189, 2022. [13] X. Li, Z. Wang, and H. Chen, "Vision transformer for gender recognition: A new perspective," IEEE Trans. Image Process., vol. 32, pp. 1856-1869, 2023. [14] S. Park and J. Kim, "EfficientNet optimization for real-time gender recognition," Pattern Recognit. Lett., vol. 168, pp. 41-48, 2023. [15] M. Rodriguez, J. Garcia, and R. Lopez, "Multi-task learning framework for gender and age estimation," Comput. Vis. Image Understand., vol. 227, Art. no. 103594, 2023. [16] R. Thompson, M. Wilson, and K. Davis, "Lightweight EfficientNetV2 for gender recognition," Neural Netw., vol. 157, pp. 28-37, 2023. [17] C. Martinez and D. Garcia, "Feature fusion techniques for robust gender recognition," IEEE Trans. Pattern Anal. Mach. Intell., vol. 45, no. 8, pp. 9234-9246, 2023. [18] J. Wilson, A. Brown, and R. Smith, "Ensemble methods for improved gender recognition," Pattern Recognit., vol. 136, Art. no. 109285, 2023. [19] Y. Zhao, X. Liu, and H. Wang, "Self-attention mechanisms in mobile networks for gender recognition," IEEE Access, vol. 11, pp. 54321-54334, 2023. [20] S. Kim and J. Lee, "Hybrid EfficientNet-Transformer architecture for gender recognition," Neural Comput. Appl., Early Access, pp. 1-12, 2024.