Blink completeness assessment using CNN–based feature extraction and transformer temporal modeling

Extended digital screen exposure has contributed to an increase in computer vision syndrome (CVS) and dry eye disease, which has increased interest in accurate eye blink analysis to support effective diagnosis and intervention. This paper presents Eye-TRANS, a hybrid deep learning architecture specifically developed for both blink detection and blink completeness assessment. Eye-TRANS combines the strengths of convolutional neural networks to extract rich spatial features from eye images with a transformer encoder that employs local self-attention and a sliding window approach to capture blink dynamics within their immediate temporal context. The local attention restricts interactions to neighboring frames, while the sliding window provides temporal context for each prediction. The resulting frame-level predictions are then aggregated to assess blink completeness, a measurement relevant to CVS research. In descriptive comparisons with the published estimates included in this study, Eye-TRANS achieves the highest blink-detection and incomplete-blink F1-scores on three of four datasets and remains competitive on the fourth. Our findings support the use of localized temporal context for blink analysis.