High-Automatical and High-Accurate Pupil Location Neural Network via FRST FPL

Pupil location refers to the location of the pupil or its center in an image. To solve the problem that the pupil location method is difficult to achieve high automation and high accuracy at the same time, this paper proposes a method combining image processing and statistical learning. In this paper, an improved algorithm of the fast radial symmetry transform (FRST) based on pupil location is proposed, namely FRSTFPL (fast radial symmetry transform for pupil location), which is used to coarsely localize the pupil in the image, followed by a shallow CNN to achieve precise localization. In addition, we construct a dataset based on the CASIA-IrisV4 iris image database and then conduct a variety of experiments. The results show that the location error of the proposed method in an image with a size of 640 × 480 pixels is 8.51 pixels, which exceeds the performance of the comparing methods. In our method, not only accurate radius and complex network are unnecessary, but also highly automated, low computational complexity, and relatively high localizing accuracy can be achieved together.