Text-Guided Feature Completion in the CLIP Embedding Space for Occlusion-Robust Facial Expression Recognition

Camera-based affective sensing degrades sharply when the face is partially occluded by hands, masks, eyeglasses or headwear. Existing occlusion-robust facial expression recognition (FER) methods either discard the information carried by occluded regions through attention re-weighting, hallucinate it back at the pixel level, or restore it in a purely visual latent space; none of them can consult an external semantic description of what the hidden region ought to look like. We propose TeCoRe, a text-conditioned completion-and-recognition framework that restores occluded patch tokens directly inside the joint vision–language embedding space of a frozen CLIP encoder. A lightweight visibility gate estimates a per-patch occlusion probability; a two-layer transformer decoder reconstructs the occluded tokens using class-level textual emotion embeddings, produced by the frozen CLIP text encoder, as both cross-attention memory and an alignment target; and a gated blend passes clean tokens through unchanged so that completion cannot corrupt visible evidence. TeCoRe is trained as supervised FER using the ground-truth emotion labels for classification and text alignment, together with an auxiliary self-supervised reconstruction objective. Because the occlusion masks are synthesised during training, the pre-occlusion CLIP tokens of the same image provide the reconstruction target without requiring additional manual annotation. The CLIP backbone is completely frozen—no low-rank adapters, no landmarks, no segmentation priors—and only about 27 M head and decoder parameters are trained. On the standard RAN occlusion and pose subsets, TeCoRe reaches 91.42% on Occlusion-RAF-DB, 84.79% on Occlusion-FERPlus, and 62.42% on Occlusion-AffectNet. It achieves the highest accuracy in our comparison on Occlusion-RAF-DB and on the RAF-DB and FERPlus pose subsets, while remaining competitive but not leading on the FERPlus and AffectNet occlusion subsets. The ablation and control studies are reported as three-seed means so that the small differences between variants are not read off a single initialisation. A backbone-generality control shows positive mean gains from the proposed modules across all three evaluated frozen CLIP configurations, with the largest gain observed in the main ViT-L/14@336 setting. These results indicate that performing occluded-token completion in a joint vision–language space, with class-level text embeddings providing shared conditioning to the completion process, is an effective and parameter-efficient route to occlusion-robust FER.