MAL-Net: Multiscale Attention Link Network for accurate eye center detection

In recent years, due to wide potential applications like human-computer interactions, eye center detection has received growing research interest. Although numerous approaches have been proposed, achieving high accuracy in the wild remains challenging because of the variations in appearance, shape, and illumination. In this paper, we formulate the eye center detection problem as an approximate segmentation task. We propose a novel method to detect pupil positions from a face image using a deep neural network with synthetic data. In particular, we introduce a deep neural network named Multiscale-Attention-Link Network (MAL-Net), where we design a Link Attention Module (LAM) and a novel Multiscale Link Structure (MLS) for accurate and robust eye center detection. Besides, a weighted loss is proposed to make the deep model pay more attention to eye centers during training. Furthermore, to address the problem of insufficient training data with enough variations in eye shape and illumination, we propose a GAN-based method named shape-GAN to generate synthetic eye images with various shapes for training. The proposed MAL-Net is evaluated on widely-used benchmarks such as TFV, GI4E, and BioID. The results demonstrate that our proposed method outperforms state-of-the-art methods for eye center detection.