Depth information plays an important role in the fields of robotics and autonomous driving. The depth map obtained by the depth sensor is relatively sparse. Researchers have proposed a large number of methods to complement the missing depth values. However, most of the existing methods aim at opaque objects. Based on the powerful representation ability of convolution neural network, this paper designed a dual-branch-guided encoder-decoder structure network. Through mask-guided branch for transparent objects, it improves the ability of the network to extract feature information of transparent objects. And spectral residual blocks improves the stability of network in training process and the ability to obtain object structure information. In addition, attention mechanism is added to improve the feature modeling ability of network space and semantic information. The network achieves state-of-the-art results on all two datasets.