Facial Emotion Recognition (FER) is an important component of intelligent human–computer interaction and affective computing because it allows machines to learn and react to human emotional conditions. All current Emotion Recognition (ER) methods are vulnerable to several drawbacks, including occlusions, lighting differences, and lack of modeling to complex spatial relations among faces. To overcome these limitations, this research proposes a Graph Transformer (GT)-based emotion classification that effectively captures relational. The collected video data undergoes frame extraction and preprocessing to improve visual quality and consistency. The faces are detected using an optimized You Only Look Once version 11 (YOLO v11) model enhanced with the Oppositional Learning-Crested Porcupine Optimization Algorithm (OL-CPOA) for improved localization accuracy and convergence. Deep facial features are extracted using a Deep Convolutional Neural Network (DCNN) integrated with Uncertainty-Aware Self-Attention (UASA) to handle noise and uncertain representations in feature maps. Further, Rectangle-Window Cross-Attention (RWCA) is employed to capture multi-scale contextual dependencies between facial regions before feeding the refined features into the GT for final emotion classification. The experimental evaluation demonstrates that the proposed model achieves a success rate of 96%, a Positive Predictive Value (PPV) of 96.14%, a sensitivity of 96.05%, and an F1-score of 96%, indicating strong classification capability.
Psoriasis is a chronic inflammatory skin disease that affects millions of people worldwide. Recently, various deep learning methods have been introduced for classifying the severity of psoriasis and segmenting affected the areas. Actual assessments necessitate the concurrent execution of multiple visual tasks, including disease detection and lesion severity evaluation. We proposed a multi-task network along with the hierarchical feature fusion attention module (HFF-AM) that extracted and integrated both local and global features of diseases from hierarchical layers. To augment the diversity of diseases, we generated masked disease images and concurrently executed three tasks—severity classification, disease segmentation, and reconstruction of masked regions using limited disease information. We investigated the interrelationships among the three tasks, thereby proposing multi-loss aggregation designs. In a comparative analysis of the state-of-the-art methods for both single-task and multi-task, the proposed method exhibited superior performance in classification and segmentation. Additionally, the proposed method achieved the best performance in the process of determining the final average loss value in a scenario that aggregated loss values by grouping classification and reconstruction tasks based on a segmentation task. This was evidenced by a severity classification accuracy of 81.5%, a disease segmentation dice similarity coefficient value of 0.938, and a reconstruction loss value of 0.034. This results indicate that the proposed MTAN achieved competitive performance across all three tasks with the corresponding single-task models for three tasks. The proposed method demonstrated a strong capability to extract lesion features despite the constraints of limited disease information, facilitating both classification and segmentation tasks. The proposed method is applicable for the evaluation of a range of skin diseases that exhibit characteristics similar to those of psoraisis.
