Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A deep convolutional neural network for salt-and-pepper noise removal using selective convolutional blocks

Ahmad Ali Rafiee, Mahmoud Farhang

TL;DR

This work targets salt-and-pepper noise removal in grayscale and color images, focusing on very high SAP densities where traditional CNNs struggle. It introduces SeConvNet, a CNN that begins with selective convolutional blocks to estimate pure SAP pixels using only non-noisy neighbors, followed by conventional CNN layers for refinement. The SeConv blocks preserve clean pixels while initializing denoising, leading to superior PSNR and SSIM across BSD68, CBSD68, and a 20-image grayscale set, especially at high densities up to 95%. The approach yields edge-preserving restorations and outperforms multiple baselines, with open-source code available for practical use in impulse-noise imaging tasks.

Abstract

In recent years, there has been an unprecedented upsurge in applying deep learning approaches, specifically convolutional neural networks (CNNs), to solve image denoising problems, owing to their superior performance. However, CNNs mostly rely on Gaussian noise, and there is a conspicuous lack of exploiting CNNs for salt-and-pepper (SAP) noise reduction. In this paper, we proposed a deep CNN model, namely SeConvNet, to suppress SAP noise in gray-scale and color images. To meet this objective, we introduce a new selective convolutional (SeConv) block. SeConvNet is compared to state-of-the-art SAP denoising methods using extensive experiments on various common datasets. The results illustrate that the proposed SeConvNet model effectively restores images corrupted by SAP noise and surpasses all its counterparts at both quantitative criteria and visual effects, especially at high and very high noise densities.

A deep convolutional neural network for salt-and-pepper noise removal using selective convolutional blocks

TL;DR

This work targets salt-and-pepper noise removal in grayscale and color images, focusing on very high SAP densities where traditional CNNs struggle. It introduces SeConvNet, a CNN that begins with selective convolutional blocks to estimate pure SAP pixels using only non-noisy neighbors, followed by conventional CNN layers for refinement. The SeConv blocks preserve clean pixels while initializing denoising, leading to superior PSNR and SSIM across BSD68, CBSD68, and a 20-image grayscale set, especially at high densities up to 95%. The approach yields edge-preserving restorations and outperforms multiple baselines, with open-source code available for practical use in impulse-noise imaging tasks.

Abstract

In recent years, there has been an unprecedented upsurge in applying deep learning approaches, specifically convolutional neural networks (CNNs), to solve image denoising problems, owing to their superior performance. However, CNNs mostly rely on Gaussian noise, and there is a conspicuous lack of exploiting CNNs for salt-and-pepper (SAP) noise reduction. In this paper, we proposed a deep CNN model, namely SeConvNet, to suppress SAP noise in gray-scale and color images. To meet this objective, we introduce a new selective convolutional (SeConv) block. SeConvNet is compared to state-of-the-art SAP denoising methods using extensive experiments on various common datasets. The results illustrate that the proposed SeConvNet model effectively restores images corrupted by SAP noise and surpasses all its counterparts at both quantitative criteria and visual effects, especially at high and very high noise densities.
Paper Structure (11 sections, 11 equations, 12 figures, 4 tables, 1 algorithm)

This paper contains 11 sections, 11 equations, 12 figures, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. The Proposed Denoising CNN Model
  3. The Network Architecture
  4. Selective Convolutional Block
  5. Experiments
  6. Datasets
  7. Preprocessing Stage
  8. Network Training
  9. Evaluation Criteria
  10. Experiments on SAP noise Removal
  11. Conclusion

Figures (12)

  • Figure 1: The architecture of SeConvNet. Note that "SeConv Block- $s$" indicates a SeConv block of size $s$.
  • Figure 2: The BSD68 dataset.
  • Figure 3: The 20 traditional test images dataset. (a) Baboon (b) Barbara (c) Blonde Woman (d) Boat (e) Bridge (f) Cameraman (g) Dark Haired Woman (h) Einstein (i) Elaine (j) Flintstones (k) Flower (l) Hill (m) House (n) Jet Plane (o) Lake (p) Lena (q) Living Room (r) Parrot (s) Peppers (t) Pirate.
  • Figure 4: Restoration results of different methods for the Peppers image with $95\%$ SAP noise. (a) Original image (b) Noisy image (c) ARmF (d) ACmF (e) IAWMF (f) NAHAT (g) DAMRmF (h) SeConvNet.
  • Figure 5: Restoration results of different methods for the Parrot image with $95\%$ SAP noise. (a) Original image (b) Noisy image (c) ARmF (d) ACmF (e) IAWMF (f) NAHAT (g) DAMRmF (h) SeConvNet.
  • ...and 7 more figures