Convolutional Neural Network Design for Eye Detection Under Low-Illumination

Abstract

The eye is an important organ in the human body for sensing and communicating with the outside world. The development of human eye detectors is essential for applications in the computer vision field, especially under low illumination. This paper proposes a convolutional neural network to detect the position of the eye in the acquired image. This network architecture exploits the advantages of convolutional neural networks combined with the concatenated rectified linear unit (C.ReLU), inception module, and Bottleneck Attention Module (BAM) to extract feature maps. Then it uses two detectors to localize the eye area using bounding boxes. The experiment was trained, evaluated on the BioID Face and Yale Face Dataset B (YALEB) dataset. As a result, the network achieves 99.71% and 99.37% of Average Precision (AP) on YALEB and BioID Face datasets, respectively.