To suppress the high-level noise of raw images from the low-light image sensor, this paper proposes a collaborative filtering algorithm based on exact noise variance of transform domain. Firstly, the noise of low-light-level images is modeled as Poisson–Gaussian mixed noise and performed by variance stabilizing transformation (VST). Secondly, a calculation method of exact noise variance is proposed based on L1 total generalized variation (L1-TGV) regularization. Finally, the denoised images are obtained by embedding the exact noise variance into block matching and three-dimensional filtering (BM3D) algorithm to improve patch matching and shrinkage accuracy. Numerical experiments on unnaturally degraded images express that the proposed method can effectively remove high-level noise and maintain image textures. Compared with BM3D algorithm, the proposed method can improve the peak signal-to-noise ratio (PSNR) by up to 2.15 dB and the structural similarity (SSIM) by up to 0.106, respectively. Moreover, the testing of the raw low-light images confirms the best performance of vision in contrast with the other four methods.