An efficient deep sclera recognition framework with novel sclera segmentation, vessel extraction and gaze detection

Highlights：

• DeepR is a deep model for sclera recognition which compares two vessel-structure pairs taking into account their affine-transformation, with no post logic in the matching process.

• An improved version of DeepR (named dual-output), improves results by a large margin achieving AUC score of 0.98 for SBVPI dataset.

• Fast and lightweight model UNet-P, uses a trained knowledge base of selective colour information that simplifies the initial task of sclera segmentation and achieves the best balance between accuracy and efficiency for sclera segmentation.

• Vessel extraction method VESD adapts to multiple datasets easily, and identifies prominent vessels that work best for recognition. Alternatively, DeepV detects vessels directly from color images without prior segmentation.

• Deep gaze detection model achieves all-correct classifications for four gaze directions using sclera-segmented images as input and also works with vessel-segmented images.

摘要

•DeepR is a deep model for sclera recognition which compares two vessel-structure pairs taking into account their affine-transformation, with no post logic in the matching process.•An improved version of DeepR (named dual-output), improves results by a large margin achieving AUC score of 0.98 for SBVPI dataset.•Fast and lightweight model UNet-P, uses a trained knowledge base of selective colour information that simplifies the initial task of sclera segmentation and achieves the best balance between accuracy and efficiency for sclera segmentation.•Vessel extraction method VESD adapts to multiple datasets easily, and identifies prominent vessels that work best for recognition. Alternatively, DeepV detects vessels directly from color images without prior segmentation.•Deep gaze detection model achieves all-correct classifications for four gaze directions using sclera-segmented images as input and also works with vessel-segmented images.