High-Accuracy Iris Center Localization Using Mediapipe Keypoints and Xception-Based Deep Regression

The precise identification of the iris centre is essential for numerous applications, such as biometrics, telemedicine, and ocular health diagnostics. This paper presents a novel approach that combines Mediapipe’s keypoint detection for accurate eye area identification with a deep regression framework based on the Xception architecture, specifically aimed at predicting the coordinates and radius of the iris centre. The model utilizes a manually annotated dataset from the Columbia Gaze dataset, incorporating an Xception backbone with three distinct regression outputs to estimate the x and y coordinates of the iris centre and its radius. Comprehensive testing was undertaken to refine critical parameters, such as the depth of the backbone layers, selection of optimizer, and resolution of input images, with training conducted over 200 epochs via the Huber loss function. The optimal configuration—comprising a 130-layer backbone, Adam optimizer, and an input resolution of $186\times 186$ pixels—produced a mean Euclidean distance ( $\mu _{ED}$ ) of 0.736 and a standardized Euclidean distance ( $S_{ED}$ ) of 2.208 on the GI4E dataset. Upon evaluation using the BioID dataset, it attained $\mu _{ED}$ and $S_{ED}$ scores of 1.560 and 3.045, respectively. The model exhibited near real-time performance, achieving an average frame processing time of 0.056 seconds (about 17.7 frames per second) on a MacBook Air M3. These findings highlight the method’s enhanced efficacy relative to current methodologies, offering an effective and reliable alternative for real-time iris localization and analysis.