There are growing interests in using cone-beam computed tomography (CBCT) for patient treatment position setup and dose evaluation in radiation therapy. The repeated use of CBCT during the course of a treatment has raised concerns of extra radiation dose delivered to patients. One way to reduce radiation dose delivered to patients during CBCT procedure is to acquire CT projection data with a lower mAs level. However, the image quality of the projection image and the reconstructed CBCT image will degrade due to excessive quantum noise as a result of low mAs protocol. In this work, we first studied the noise properties of CBCT projection data from repeated scan and then improved low-dose CBCT image quality by restoring CBCT projection images based on an improved noise model of CBCT projection data. Analysis of repeated measurements show that noise is correlated among nearest neighbors in projection data, i.e., covariance matrix of projection data noise is non-diagonal. The covariance matrix of noise provides the knowledge of second-order statistics of noise, which may lead to more accurate estimation for statistical image reconstruction and restoration algorithm. We constructed the penalized weighted least-squares (PWLS) objective function by incorporating the noise correlation of CBCT projection data. The optimal solution of the line integrals is then estimated by minimizing the PWLS objective function. A quality assurance phantom was used to evaluate the presented algorithm for noise reduction in low-dose CBCT.