A Comprehensive Face Parsing Framework for Anxiety Detection using Deep Learning

The increasing number of adolescents suffering from anxiety highlights the need for Anxiety Detection systems. Although there are some conventional systems based on face modality, there is a compelling need for an end-to-end framework with a focus on early detection of anxiety. Although some works have presented deep learning-based multichannel approaches, they lack investigation of the impact of channels on model performance. Furthermore, semi-occluded faces are an important trait in real-world situations. However, several works suffer from lack of efficient solution. To overcome these challenges, we propose a novel multimodal Face Parsing-based system, the Cohort-Based User-Centered Framework (CB-UCF). CB-UCF assesses adolescent mental health, detects early markers of anxiety, and provides recommendations. This paper presents the experimental results of the Emotion Recognition module in CB-UCF. A novel algorithm, Visage Region Segmentation-Based Face Parsing (VRS-FP) is introduced. VRS-FP encompasses an innovative penta-channel pipeline of convolutional neural networks. Unlike traditional methods that rely on a single feature vector, VRS-FP introduces five feature vectors per input to promote robustness. Various interesting obscurations are introduced in the images to validate the methodology. In addition, two custom deep models are proposed, the Single Channel CNN (SCNN) and Penta Channel CNN (PCNN). SCNN achieves a high accuracy of 96.95% while PCNN achieves an impressive accuracy of 98.61% on CK+. SCNN and PCNN achieve high accuracies of 76.47% and 78.86% respectively on FER2013. The proposed work demonstrates a marked increase in performance over the existing literature. The promising results and contributions of this work will significantly advance research.