PECSA Net: Pupil Segmentation Using Attention Integrated Encoder-based Multi-Level Feature Extractor for Analysis of Ocular Dynamics

Abstract

Pupillometry is considered as an emerging technique for analysing the dynamic variations of pupil structure non-invasively. Delineation of the pupil region from ocular images plays a crucial role in the prediction of microexpressions and subtle variations during the analysis. Occlusions, deformations, and motion artifacts are more prevalent during the capture of micro-ocular expressions and often resulting in a suboptimal pupil mask with minimal structural similarity. This proposed study introduces a Swin incorporated modified encoder-decoder paradigm, “PECSA Net” for reliable segmentation of the pupil region. The local features including boundaries and edges, are captured through convolutional layers in the PECSA Net. Moreover, the contextual features including the overall spatial relationship within the ocular images, are captured using the deeper layers consist of an attention-enhanced multiscale contextual information module. Experimental results on CASIA dataset shows that the framework achieves superior performance in occluded conditions compared to conventional neural network-based approaches with a dice score of 0.97. Further, the model is validated with a multifaceted eye tracking dataset captured using high speed camera to evaluate the generalization capability, achieving a dice score of 0.79. Extensive architecture-based ablation studies are also carried out to identify the significance of each component of PECSA Net. Thus, by integrating attention enhanced multi-level feature extractor with an encoder-decoder paradigm, the proposed study provides a robust segmentation under challenging conditions, enabling reliable extraction of features for analyzing pupil dynamics.