基于空洞全卷积网络的非理想状态的瞳孔定位

Non-ideal pupil positioning based on hole fully convolutional network

针对不良光照、眼镜和隐形眼镜导致的大量反射、瞳孔位于图像边缘等非理想状态下的瞳孔定位的情况,采用了一种基于空洞全卷积网络非理想状态的瞳孔定位方法。该网络是端到端系统,由两个完全卷积网络组成。首先,输入非理想状态的瞳孔图像,经过下采样和空洞卷积模块逐渐展现环境信息,随后在上采样过程中,结合了下采样的各层信息和上采样的输入信息来还原细节信息,输出仅含有瞳孔部分的圆形区域的白色背景图。最终使用椭圆拟合算法,对图像的瞳孔进行拟合,得到瞳孔的中心坐标值。由于神经网络去掉了主要的干扰因素,仅保留目标对象圆形区域,因此,该算法节省了人工成本,解决了阈值调整对处于非理想状态的瞳孔图像不灵敏的问题,并且实验结果表明,该方法在公开的二十四个数据集的平均检测率达到91.7%。

Aiming at the pupil positioning in non-ideal conditions such as poor lighting,large reflections caused by glasses and contact lenses,and pupils at the edge of the image,a non-ideal pupil positioning method based on a hollow full convolutional network is adopted.The network is based on an end-to-end system and consists of two fully convolutional networks.First,input the non-ideal pupil image,and gradually display the environmental information through the down-sampling and hole convolution module.Then,in the up-sampling process,the down-sampled layers of information and the up-sampled input information are combined to restore the detailed information.Then,the white background image with only the circular area of the pupil is output.Finally,the ellipse fitting algorithm is used to fit the pupil of the image to obtain the center coordinates of the pupil.Because the neural network removes the main inter-ference factors and only retains the circular area of the target object,the labor cost is subtracted,and the problem that the threshold adjustment is not sensitive to the pupil image in a non-ideal state is solved.Experiment has shown that the average detection rate of this method in the twenty-four public datasets reaches 91.7%.