Optimized Deep Convolutional Neural Network for Robust Occluded Facial Expression Recognition

Occluded facial expression recognition (OFER) poses a formidable challenge in real-world applications, particularly in human-computer interaction and affective computing. Despite recent advancements, existing methodologies often struggle to maintain optimal accuracy under occlusion constraints. This study proposes a novel hybrid framework that synergizes handcrafted and deep learning-based features to enhance robustness and precision in emotion recognition. Specifically, we integrate Histogram of Oriented Gradients (HoG), facial landmark descriptors, and sliding window-based HoG representations with deep convolutional neural network (CNN) features, leveraging their complementary strengths. Our experimental design explores multiple feature fusion strategies, including CNN-based automated classification and a hybrid model incorporating Dlib-extracted landmarks with HoG-CNN integration. Comparative analysis against state-of-the-art approaches demonstrates that our multi-feature fusion technique significantly improves recognition accuracy, achieving a remarkable 96% accuracy on benchmark datasets such as RAF-DB and AffectNet. However, we observe a marginal decline in performance with increased dataset complexity, emphasizing the need for scalable solutions. This research underscores the efficacy of integrating handcrafted and deep learning-driven features, offering a promising direction for advancing occlusion-robust facial expression recognition in dynamic environments.