The accuracy of visibility detection greatly affects daily life and traffic safety. Existing visibility detection methods based on deep learning rely on massive haze images to train neural networks to obtain detection models, which are prone to overfit in dealing with small samples cases. In order to overcome this limitation, a large amount of measured data are used to train and optimize the convolutional neural network, and an improved DiracNet method is proposed to improve the accuracy of the algorithm. On this foundation, combined multi-mode algorithm is proposed to achieve small samples fitting and train an effective model in a short time. In this paper, the proposed improved DiracNet and the combined multi-mode algorithm are verified by using the measured atmospheric fine particle concentration data (pm1.0, pm2.5, pm10) and haze video data. The validation results demonstrate the effectiveness of the proposed algorithm.

能见度检测的准确性极大地影响着日常生活和交通安全。现有基于深度学习的能见度检测方法依靠海量雾霾图像训练神经网络获得检测模型，在处理小样本情况时容易出现过拟合。为了克服这一限制，利用大量实测数据对卷积神经网络进行训练和优化，提出了一种改进的DiracNet方法来提高算法的精度。在此基础上，提出组合多模式算法，实现小样本拟合，并在短时间内训练出有效的模型。本文利用实测大气细颗粒物浓度数据（pm1.0、pm2.5、pm10）和雾霾视频数据验证了提出的改进DiracNet和组合多模算法。