Tomographic synthetic aperture radar (TomoSAR) is an advanced SAR interferometric technique to retrieve 3-D spatial information. However, decorrelation effects degrade the quality of interferometric phases, resulting in errors in the reconstruction. In this paper, we propose a denoising method based on the unsupervised convolution neural network (CNN) with a loss function combining the deterministic descriptive regularization and total variation (TV) term. It can improve both the accuracy and completeness of the reconstructed 3-D point clouds, which is verified by experiments on simulated and real SAR images.