A Vision Transformer Framework for Explainable Facial Affect Recognition Toward Assistive Mental State Screening

Automated Facial Emotion Recognition (FER) provides an objective, non-invasive computational modality for preliminary behavioral assessment and affective monitoring. However, conventional Convolutional Neural Networks (CNNs) commonly possess limited capacity to capture wide-ranging semantic correlations among distant facial landmarks, and their opaque decision-making mechanisms impede clinical trustworthiness. In this work, we present an end-to-end explainable affective computing framework which integrates a fine-tuned Vision Transformer (ViT-B/16) backbone with Gradient-weighted Class Activation Mapping (Grad-CAM). The proposed architecture partitions aligned facial images into non-overlapping spatial patches, projects them into a continuous latent space with positional encodings, and models global self-attention across seven primary affective expressions: anger, disgust, fear, happiness, sadness, surprise, and neutral. To overcome the lack of inherent 2D spatial feature encodings in standard transformer encoders, we introduce a 1D-to-2D spatial token reshaping module that extracts gradients at the penultimate LayerNorm stage to compute localized saliency maps. The prototype demonstrates high diagnostic discriminability on established benchmark data as sustaining an inference latency under 85 ms on edge GPU accelerators. By visually correlating active attention heatmaps with particular Action Units (AUs) of the Facial Action Coding System, our system delivers verifiable, interpretable visual evidence suitable for assistive clinical processes.