Improving Facial Expression Recognition in real-world Environments

Facial expressions serve as fundamental cues for understanding human emotions and are a key component of affective computing. Recent advances in deep learning, especially Convolutional Neural Networks (CNNs), have made automated emotion recognition increasingly accurate and scalable. This paper introduces DCRNet, a hybrid deep neural network architecture designed to improve Facial Expression Recognition (FER) under real-world conditions such as occlusion, pose variation, and lighting inconsistency. The network integrates a pre-trained DenseNet121 backbone, multiple Convolutional Block Attention Modules (CBAM), and residual connections to enhance discriminative learning and gradient flow. Preprocessing employs adaptive gamma correction and facial landmark localization, ensuring optimal photometric normalization and emphasis on expressive regions of the face. Comprehensive experiments demonstrate that DCRNet achieves accuracies of 65.80%, 98.98% and 96.25% on the AffectNet, CK+, and KDEF datasets, respectively. It outperforms several recent FER models while maintaining a compact footprint of 11.6 million parameters. Cross-validation across different datasets confirms strong generalization. Statistical significance testing (McNemar and bootstrap analysis) verifies that performance gains are not due to random initialization. Further evaluation includes inference latency, FLOPs, and energy usage on GPU and ARM devices, confirming suitability for edge deployment. Finally, ethical and bias considerations are discussed to ensure responsible use in healthcare, education, and human-machine interaction.