Source-linked AI summary
Enhanced CNN for image denoising
Chunwei Tian, Yong Xu, Lunke Fei, Junqian Wang, Jie Wen, Nan Luo
TL;DR
Image denoising CNNs can be difficult to train and may encounter gradient and computational challenges as networks become deeper. The paper proposes ECNDNet, combining residual learning, batch normalization, and dilated convolutions, and reports stronger denoising performance than popular methods.
Problem
Deep image-denoising networks face training difficulties, vanishing or exploding gradients, and computational costs associated with increasing depth.
Method
ECNDNet combines residual learning, batch normalization, and dilated convolutions to improve trainability, convergence, context capture, and computational efficiency.
Results
ECNDNet is reported to outperform popular denoising methods including FFDNet, IRCNN, and BM3D.
Takeaways & Limitations
The reported results support ECNDNet as an effective deep CNN approach for image denoising.
Abstract
from arXiv · showhide
Owing to flexible architectures of deep convolutional neural networks (CNNs), CNNs are successfully used for image denoising. However, they suffer from the following drawbacks: (i) deep network architecture is very difficult to train. (ii) Deeper networks face the challenge of performance saturation. In this study, the authors propose a novel method called enhanced convolutional neural denoising network (ECNDNet). Specifically, they use residual learning and batch normalisation techniques to address the problem of training difficulties and accelerate the convergence of the network. In addition, dilated convolutions are used in the proposed network to enlarge the context information and reduce the computational cost. Extensive experiments demonstrate that the ECNDNet outperforms the state-of-the-art methods for image denoising.
1 Introduction
Image denoising methods face manual parameter tuning, complex optimization, and deep-network training problems. ECNDNet addresses these issues with residual learning, batch normalization, and dilated convolutions, and experiments report stronger performance than popular alternatives.
- Traditional denoising methods require manually tuned parameters and complex optimization, increasing computational cost.
- Deep denoising networks can suffer vanishing or exploding gradients and increased computational cost as depth grows.
- ECNDNet combines residual learning, batch normalization, and dilated convolutions to improve trainability, convergence, context capture, and computational efficiency.
- Extensive experiments report that ECNDNet outperforms FFDNet, IRCNN, and BM3D.
- The proposed network is 17 layers deep, which the authors report can reduce computational cost.
2 Related work
The paper uses residual learning and batch normalization to address deep-network optimization problems, while dilated convolutions enlarge receptive fields without the same parameter and depth increases. ECNDNet combines these components in its architecture.
- BN and residual learning: Deep networks may experience vanishing or exploding gradients and difficulty training as their depth increases.
- BN and residual learning: Residual learning adds the input x to the stacked-layer output f(x), producing f(x) + x for the next layer.
- Dilated convolution: Increasing width enlarges receptive fields but can add parameters, overfit, and increase computational cost; increasing depth can create gradient problems.
- Dilated convolution: Dilated convolution enlarges the receptive field through a dilation factor while balancing feature capture and computational cost.
3 Proposed denoising method
ECNDNet combines residual learning, batch normalisation, and dilated convolution in a 17-layer denoising network designed to ease training, enlarge context, and reduce computational cost.
- Training objective: The denoising objective trains f(y) using noisy image patches yj and corresponding label patches xj.Image patches are used to reduce computational cost.
- Network architecture: ECNDNet consists of dilated convolution, residual learning, batch normalisation, convolution, and ReLU.Dilated convolutions are placed at the 2nd, 5th, 9th, and 12th layers.
- Design rationale: The proposed network uses 17 layers, residual learning, and batch normalisation to address gradient problems and improve training efficiency.Residual learning is intended to prevent vanishing or exploding gradients, while batch normalisation accelerates convergence.
- Design rationale: Dilated convolutions capture more context while reducing the computational cost of the designed network.This design choice is presented as complementary to residual learning and batch normalisation for image denoising.
- Results: For one grayscale image at noise level 15.07 dB, ECNDNet reaches 23.19 dB, compared with 22.56 dB for BM3D and 22.83 dB for WNNM.The listed example also reports 22.81 dB for EPLL.
- Results: For another grayscale image at noise level 24.59 dB, ECNDNet reaches 33.11 dB, compared with 31.85 dB for BM3D and 32.71 dB for WNNM.The listed example also reports 32.10 dB for EPLL.
4 Experimental results
The experiments train a 17-layer ECNDNet and evaluate it on BSD68 and Set12 against established denoising methods. Results are reported using PSNR and runtime, with ECNDNet showing strong denoising performance across image categories.
- Experimental setting: ECNDNet is trained as a 17-layer network using Adam, with 180 training epochs and batch size 128.The learning rate decreases from 1×10−3 to 1×10−8 across training.
- Datasets and baselines: The evaluation uses 400 180 × 180 PNG images for Gaussian denoising, with BSD68 and Set12 as test datasets.Comparisons include BM3D, WNNM, EPLL, CSF, TNRD, IRCNN, and multilayer perceptron methods.
- Evaluation measures: Performance comparisons include image sizes of 256 × 256, 512 × 512, and 1024 × 1024 at noise level 25.Table 3 reports runtime for different methods under these settings.
- Denoising results: ECNDNet achieves good performance for each image category in the Set12 evaluation.Table 2 marks the best PSNR in italic and the second-best PSNR in bold.
- Denoising results: 30.39 dB is the reported average PSNR difference between ECNDNet and BM3D at noise level 25.The passage states that ECNDNet’s average PSNR is 30.39 dB higher than BM3D.
5 Run time
The runtime experiment compares ECNDNet and DnCNN-s for gray-image denoising across three image sizes at noise level 50. Runtime is treated alongside PSNR as an important evaluation factor.
- Runtime evaluation: Runtime is tested for 256 × 256, 512 × 512, and 1024 × 1024 noisy images with s = 50.PyTorch is used to measure the processing time of DnCNN-s and ECNDNet.
- Evaluation criteria: PSNR and image-processing runtime are identified as two important factors for evaluating image denoising methods.The runtime experiment specifically concerns gray-image denoising.
6 Conclusion
The paper concludes that ECNDNet combines batch normalisation, residual learning, and dilated convolution for image denoising. Experiments show greater effectiveness than popular denoising methods, while future work targets real noisy images.
- Conclusion: ECNDNet is proposed as a deep CNN for solving the image denoising problem.The conclusion presents ECNDNet as the paper’s central network.
- Conclusion: Batch normalisation, residual learning, and dilated convolution are combined to enhance ECNDNet’s denoising performance.The components address training difficulty, gradient problems, context extraction, and computational cost.
- Conclusion: ECNDNet is reported as more effective than popular denoising methods such as IRCNN.The conclusion attributes this result to extensive experiments.
- Future work: Future work will combine model-based optimisation and discriminative learning methods for real noisy images.This scope is stated as a future direction rather than a demonstrated result.