Search

In this paper, we propose the incorporation of noisy image patches and the impulse response of a low-pass filter (LPF) in a convolutional neural network (CNN) to denoise Poisson--Gaussian noise in low-dose computed tomography (LDCT) images. The approach is referred to as fast and flexible denoising CNN (FFDNet)-impulse response (FFDNet-IR) in this paper. The power spectrum sparsity LPF (SLPF) allows low-frequency components to pass through while suppressing higher frequency components by the sparsity approach of the power spectrum, and it is employed to determine the impulse response of LPF. Three well-known types of LPF, namely, Direct LPF, Gaussian LPF, and Butterworth LPF, are also considered to obtain the impulse response of LPF. In the FFDNet-IR, both the noisy image patches and the IR of the LPF are sequentially inputted into the FFDNet to eliminate the Poisson--Gaussian noise. This approach enhances the denoising performance in LDCT images compared with the conventional FFDNet in the evaluation metrics of the peak signal-to-noise ratio (PSNR), structural similarity (SSIM), and feature similarity (FSIM). Moreover, the FFDNet-IR trained with the Poisson--Gaussian noise model demonstrates the generalization ability and effectively eliminates only Poisson or Gaussian noise. The experiments indicate that the FFDNet-IR more effectively suppresses the noise artifacts and preserves image details compared with the baseline FFDNet, as well as traditional methods such as block-matching and 3D filtering (BM3D) and nonlocal mean (NLM) for LDCT image denoising. [ABSTRACT FROM AUTHOR]

In this paper, we propose the sequential input of a noisy image patch and the impulse response of a low-pass filter (LPF) in the training of the conventional fast and flexible solution for CNN-based image denoising (FFDNet) architecture, which enhances denoising performance and edge preservation and achieves high perceptual quality. The proposed method consists of two steps. In the first step, the power spectrum sparsity is utilized to determine the impulse response of LPF and the resulting impulse response is added to the noisy image patch in a sequential form to estimate the low- and high-frequency components of the input image. In this step, the use of three different types of LPF is also considered. In the second step, the FFDNet architecture, a deep-learning-based image denoiser, is employed. The proposed method achieves satisfactory denoising performance for grayscale and color datasets on synthetic additive white Gaussian noise (AWGN) in terms of the peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), feature similarity index (FSIM), and learned perceptual image patch similarity (LPIPS) compared with the original FFDNet. The performances on realistic noise and for chest X-ray images are also investigated. [ABSTRACT FROM AUTHOR]