Abstract: Aiming at  the problem that the traditional iris location algorithm was difficult to achieve accurate location, resulting in unstable recognition effect, we proposed an iris location algorithm based on improved YoloV4 network. Firstly, YoloV4 combined with MobileNetV3 was used for rough positioning of the inner and outer circles of the iris, and then the gray difference between the pupil, iris and sclera was used for fine positioning of the inner and outer circles of the iris. At the same time, K-means++ clustering algorithm was used to generate a priori box. The fast soft-DIoU-NMS algorithm was used to remove the redundant boxes in the prediction process and improve the detection rate of the algorithm. The Focal Loss was used as a category loss function. The comparative experimental results show that the algorithm runs faster, the location accuracy is higher, and the effect of recognition algorithm is more obvious.

Key words: iris location, YoloV4, fast soft-DIoU-NMS algorithm, K-means++ clustering