Abstract: In order to improve the accuracy of ground-based cloud image segmentation, a symmetrical densely connected convolution neural network based cloud image segmentation method is proposed for ground-based cloud image segmentation. The proposed new network structure firstly extracts the features of ground-based cloud images through ordinary convolution layer, then further processes the feature images through continuous dense connection blocks and upper sampling modules, and finally fuses the shallow and deep features of the network in parallel to achieve accurate segmentation of the ground-based cloud images. In the dense block, cross-layer connection is used to transfer the features of the layer used in the network, which makes the features of cloud images reused, and at the same time reduces the problem of gradient disappearance in the training process to a certain extent. The feature maps of shallow and deep networks are connected in parallel to achieve further accurate segmentation of the ground-based cloud images. The experimental results show that this method can improve the segmentation accuracy and has good generalization effect compared with other machine learning methods for ground-based cloud segmentation.

Key words: deep learning, symmetric convolutional neural network with dense connection, segmentation of image, ground-based cloud images

摘要： 为了提高地基云图分割的精度，提出一种对称式密集连接卷积神经网络的云图分割方法进行地基云图分割研究。提出的新的网络结构通过普通卷积层提取地基云图特征，通过连续的密集连接块和上采样模块对特征图进一步处理，通过并联方式融合网络浅层和网络深层的特征图从而实现对地基云图精确的分割。其中，密集块中采用跨层连接的方式实现了网络中所用层的特征传递，使得云图特征得到复用，同时一定程度上减轻了训练过程中的梯度消失问题，通过并联浅层网络和深层网络的特征图实现了对地基云图的进一步精确分割。实验结果表明，该方法与其他用于地基云图分割的机器学习方法相比，能够提高地基云图的分割准确率，具有良好的泛化效果。

关键词: 深度学习, 对称式密集连接卷积神经网络, 图像分割, 地基云图