Automated segmentation of palpebral fissures from eye videography using a texture fusion neural network

Automated segmentation of palpebral fissures depicted on images of eye videography plays a crucial role in identifying blinks (especially incomplete blinks) and detecting some eye disorders (e.g., dry or tired eyes). To this end, we propose a texture fusion segmentation network (called TF-Net) on the basis of the popular U-Net for automatically segmenting palpebral fissures from eye videography. The proposed network is constructed by defining two novel subtraction convolutional blocks and introducing them into the U-Net. These subtraction blocks are able to suppress motion artifacts and complex background in some degrees and enhance the sensitivity of the proposed network to objects of interest, as compared to the U-Net. Extensive experiments on our collected blink images from eye videography and three public datasets demonstrated that the proposed network obtained, on average, the Dice score (DS), Jaccard similarity (JS), and Hausdorff distance (HD) of 0.9445, 0.9099, and 5.3657, respectively, on the blink image dataset, and 0.8364, 0.7434, and 4.5745 for 14 objects of interest on three public datasets, which surpassed the U-Net and its multiple variants.