Abstract: Refractive error is a very common and highly detrimental ophthalmic problem to the development of visual function. Accurate and convenient refractive detection techniques are of great importance for timely detection of refractive error problems and for adopting corresponding measures for intervention. Although currently refractive screening device can quickly detect refraction, there are two main problems：the detection accuracy is low, and the requirements for the degree of cooperation of the tested people are high. Therefore, this paper proposes a new refractive detection method, which obtains face NIR images using an optical system based on the principle of eccentric photographic refraction, processes face NIR images using image processing technology to obtain left and right pupil images and pupil position information. Then, using the mixed data multi-input neural network model proposed in this paper that combines the depth separable convolution and SE module for training and diopter calculation. Compared with refractive detection methods, which are based on the principle of eccentric photographic refraction, this method is expected to achieve higher accuracy with the expansion of the data set, and this method, which uses pupillary position information as input to the model, can solve the problems of traditional algorithms that require a higher degree of cooperation from the tested people. This article is a useful exploration for the new methods of refraction detection. The use of this method is conducive to more convenient refractive screening and provides a basis for the realization of non-contact self-service refractive screening.

Key words: refractive detection, image processing;eccentric photography optometry, deep learning

摘要： 屈光不正是一种非常常见且对视功能发育有严重危害的眼科问题。准确与方便的屈光检测技术，对于及时发现屈光不正问题以及采取相应措施进行干预具有非常重要的意义。目前的屈光筛查设备虽然能较快进行屈光检测，但主要存在两个问题：检测准确度较低，对被测者配合度要求较高。因此，提出一种新的屈光检测方法，此方法使用基于偏心摄影验光原理的光学系统获取人脸面部近红外图像，使用图像处理技术对面部近红外图像进行处理，得到左右瞳孔图像和瞳孔位置信息，使用提出的结合了深度可分离卷积和SE模块的混合数据多输入神经网络模型进行训练与屈光度的计算。与传统偏心摄影验光原理的屈光检测方法相比，此方法有望随着数据集的扩增而达到更高的准确度，并且此方法将瞳孔位置信息作为模型的输入，可以解决传统算法对被测者配合度要求较高的问题。该研究是对屈光检测新方法的一种有益探索，使用此方法有利于屈光筛查更便利地进行，为实现非接触自助式的屈光筛查提供基础。

关键词: 屈光检测, 图像处理, 偏心摄影验光, 深度学习