DPW-SDNet: Dual Pixel-Wavelet Domain Deep CNNs for Soft Decoding of JPEG-Compressed Images

1. Introduction

• Compression artifacts are especially problematic at low bit-rates and can reduce object-detection and classification accuracy.The paper motivates improving compressed-image quality partly by these downstream effects.Source: §1 · p. 1
• Soft decoding, deblocking, and compression-artifact reduction are used interchangeably for JPEG-image restoration in this paper.The paper identifies soft decoding as its focus because JPEG is a commonly used still-image compression standard.Source: §1 · p. 1
• DPW-SDNet uses parallel pixel- and wavelet-domain CNNs whose estimates are combined into a final JPEG soft-decoded image.The paper describes this dual-branch design as its main contribution.Source: §1 · p. 2
• The networks use smaller tensors formed from the compressed image and its 1-level DWT coefficients, and the authors report state-of-the-art performance while maintaining efficiency.The pixel and wavelet subnetworks receive the respective reshaped inputs.Source: §1 · p. 2

2. Related Work

• Earlier soft-decoding methods are grouped into enhancement-based filtering, restoration-based methods, and learning-based methods.The paper frames JPEG soft decoding as estimating the original image from its compressed version.Source: §2 · p. 2
• Restoration-based methods use image, compression, or parameter priors, but their complex optimization can make them time-consuming.Examples include low-rank and group-sparse models, image priors, graph models, and DCT quantization constraints.Source: §2 · p. 2
• Deep-learning approaches are reported to outperform conventional soft-decoding methods in restoration performance and running time.The cited work includes CNNs, trainable nonlinear reaction-diffusion models, and dual-domain JPEG deblocking networks.Source: §2 · p. 2
• The paper's running-time comparisons refer to testing time, although deep-learning restoration methods may require time-consuming training.Training and testing time are distinguished explicitly.Source: §2 · p. 2
• The pixel-branch downsampling is reversible, and the flowchart specifies 20-layer pixel and wavelet subnetworks with 3 × 3 convolution kernels.The figure identifies the number next to each convolutional layer as its kernel count.Source: Figure 3 · p. 3, Figure 2 · p. 3
• DPW-SDNet differs from earlier deep soft-decoding algorithms by using parallel pixel- and wavelet-domain branches and two tensor inputs rather than the original image and DWT coefficients.The paper reports competitive restoration performance and execution speed; extensions to other compression standards are described as straightforward.Source: §2 · p. 3

3. Proposed DPW-SDNet

• DPW-SDNet combines parallel pixel- and wavelet-domain restoration branches, then equally weights their estimates to produce the final soft-decoded image.The branches use smaller tensors rather than the original compressed image and its DWT sub-bands as inputs.Source: §3 · p. 3, §3.3 · p. 4

3.1. The Pixel Domain Branch

• Four spatially sampled sub-images form a smaller four-channel input for P-SDNet, reducing computational complexity and enlarging the receptive field.The network predicts a four-channel residual, whose feature maps are reassembled into the pixel-domain estimate.Source: §3.1 · pp. 3–4

3.2. The Wavelet Domain Branch

• W-SDNet processes the four 1-level DWT sub-bands as a four-channel tensor and predicts wavelet-coefficient residuals for inverse transformation.Its architecture matches P-SDNet in depth, kernel count, and kernel size.Source: §3.2 · p. 4

3.3. The Combination of the Dual-Branch

• Because the two estimates come from different spaces, the method combines them by equal weighting rather than adding a learned fusion network.The stated design choice considers computational complexity.Source: §3.3 · p. 4

4.2. Soft Decoding Performance Evaluation

• The evaluation uses Classic5 and LIVE1 at QF = 10, 20, 30, and 40, with a dedicated DPW-SDNet model trained for each quality factor.The five competitors use the original codes and models provided by their authors.Source: §4.2 · pp. 4, 6
• On Classic5 and LIVE1, DPW-SDNet averages 0.30–0.34 dB PSNR, 0.0030–0.0098 SSIM, and 0.04–0.24 dB PSNR-B gains over DnCNN-3 [34], with one PSNR-B exception.It consistently outperforms five competitors except for PSNR-B on Classic5 at QF = 40, where CONCOLOR is superior.Source: Table 1 · p. 5, §4.2 · p. 6
• At QF = 10, DPW-SDNet is reported to reduce artifacts, restore details, and produce more perceptually appealing images than the compared methods.The visual comparisons include Barbara, Bike, and Lighthouse3.Source: Figure 4 · p. 5, §4.2 · p. 6, Figure 5 · p. 6

4.3. Discussion on Dual-Domain Soft Decoding

• DPW-SDNet outperforms its pixel-only P-SDNet and wavelet-only W-SDNet variants, supporting the effectiveness of combining both domains.The comparison reports average PSNR, SSIM, and PSNR-B scores on Classic5 and LIVE1.Source: §4.3 · p. 7, Table 2 · p. 7

4.4. Discussion on Blind Soft Decoding

• The universal B-DPW-SDNet is slightly inferior to dedicated DPW-SDNet but outperforms DnCNN-3 [34] in most cases, offering competitive results when QF is unknown.Dedicated models provide better restoration when QF is known; the universal model offers more flexibility.Source: §4.4 · pp. 7–8, Table 3 · p. 8

4.5. Empirical Study on Training Convergence and Running Time

• Training converges after about 200,000 iterations; the experiments use a maximum of 300,000 iterations, and one model takes about nine hours to train on a GeForce GTX 1080 Ti.The convergence curves are shown for QF = 40, with similar trends reported for other QFs.Source: §4.5 · p. 8, Figure 7 · p. 8
• DPW-SDNet takes about twice as long as either single-domain branch but less time than the other compared algorithms, and GPU use can greatly accelerate it.The runtime comparison covers three representative image sizes in Classic5 and LIVE1.Source: §4.5 · p. 8, Figure 8 · p. 8

• DPW-SDNet restores JPEG-compressed images in both pixel and wavelet spaces using deep CNNs.Source: §5 · p. 8
• Using 4-channel tensors rather than 2-dimensional images makes DPW-SDNet efficient and effective.Source: §5 · p. 8
• Benchmark-dataset experiments demonstrate the algorithm's effectiveness and efficiency; future work includes other compression standards and image restoration problems.Source: §5 · p. 8