A new nonlocal means based framework for mixed noise removal

A serial attention module‐based deep convolutional neural network for mixed Gaussian‐impulse removal

“…The average PSNR, SSIM, and FSIM measures are given in Tables 5–7, respectively. For the mixture of AWGN+SPIN, it can be seen that the proposed SACNN achieves similar results to NM‐CNN [30] and outperforms the other five algorithms. But for the mixtures of AWGN+RVIN and AWGN+SPIN+RVIN, SACNN achieved superior results to all the other algorithms.…”

“…This section describes the extensive experiments used to evaluate SACNN. For the three types of noise, we compared SACNN against six competitive algorithms, including two traditional models (ISNP [24] and LSLR [25]) and four methods based on CNN (BDCNN [29], TL‐CNN [28], VA‐CNN [27], and NM‐CNN [30]). The PSNR, SSIM, and FSIM [39] metrics were calculated to estimate the results for SACNN and the six competing algorithms.…”

“…From Tables 1 and 2, it can be seen that the proposed SACNN outperforms ISNP [24], LSLR [25], BDCNN [29], TL‐CNN [28], and VA‐CNN [27]. SACNN achieves results similar to those of NM‐CNN [30]; the advantages of SACNN against NM‐CNN are increasingly clear for larger IN. For the mixture of AWGN+SPIN, the reason why all algorithms can achieve perfect results is that IN can be detected accurately.…”

“…Figure 8(a) is the original image and Figure 8(b) depicts the image “Starfish” corrupted by AWGN+SPIN+RVIN $$(\sigma =5, p_{\rm r}=5\%, p_{\rm s}=50\%)$$. We can see that LSLR [25] and VA‐CNN [27] cannot remove all IN; ISNP [24], BDCNN [29], TL‐CNN [28], and NM‐CNN [30] blur the image texture details, while SACNN reconstructs a notably superior fine texture.…”

“…In ref. [30], a nonlocal means‐based CNN (NM‐CNN) method was presented. By utilizing nonlocal means to make the distribution of noise approximately Gaussian, the CNN was further used to remove AWGN.…”