Source-linked AI summary
Implicit Dual-domain Convolutional Network for Robust Color Image Compression Artifact Reduction
Bolun Zheng, Yaowu Chen, Xiang Tian, Fan Zhou, Xuesong Liu
TL;DR
Prior dual-domain methods struggle with color images because their compression differs from grayscale, and recovery can decline when the quantization-quality factor mismatches the pretrained model. IDCN estimates DCT-domain losses without DCT and adds position labeling, while IDCN-f targets a wide range of compression qualities. Reported ablations favor simplified options, and IDCN's LIVE1 speed is similar to simplified RDN's.
Problem
Dual-domain methods struggle with color images because color-channel compression differs, and recovery can decline when the quantization-quality factor mismatches the pretrained model.
Source: §I · pp. 1–2Method
IDCN estimates DCT-domain losses without DCT using an extractor–corrector framework, and its flexible version uses pixel-position labeling to handle varied compression qualities.
Source: §I · p. 2Results
Ablations report improvement with the simplified -SC option and possible improvement with -SL; on LIVE1, IDCN runs at a speed similar to simplified RDN and slightly slower than MemNet and S-Net.
Source: B. Ablation Investigation · p. 9, D. Comparison with State-of-the-Art… · p. 11Takeaways & Limitations
IDCN-f is reported to handle a wide range of compression qualities with one model weight.
Source: §I · p. 2
Abstract
from arXiv · showhide
Several dual-domain convolutional neural network-based methods show outstanding performance in reducing image compression artifacts. However, they suffer from handling color images because the compression processes for gray-scale and color images are completely different. Moreover, these methods train a specific model for each compression quality and require multiple models to achieve different compression qualities. To address these problems, we proposed an implicit dual-domain convolutional network (IDCN) with the pixel position labeling map and the quantization tables as inputs. Specifically, we proposed an extractor-corrector framework-based dual-domain correction unit (DCU) as the basic component to formulate the IDCN. A dense block was introduced to improve the performance of extractor in DRU. The implicit dual-domain translation allows the IDCN to handle color images with the discrete cosine transform (DCT)-domain priors. A flexible version of IDCN (IDCN-f) was developed to handle a wide range of compression qualities. Experiments for both objective and subjective evaluations on benchmark datasets show that IDCN is superior to the state-of-the-art methods and IDCN-f exhibits excellent abilities to handle a wide range of compression qualities with little performance sacrifice and demonstrates great potential for practical applications.
I. INTRODUCTION
JPEG's blockwise DCT quantization can degrade visual quality at high compression, while prior dual-domain methods struggle to use DCT priors for color restoration. IDCN estimates DCT-domain losses without an explicit DCT transform and includes a flexible version for varied compression qualities.
I. INTRODUCTION
- DCT-domain methods can cause chromatic aberrations when color channels are restored separately, and color compression differences complicate use of a single luminance-trained model.Pixel-domain learning captures inter-channel correlations but has difficulty directly introducing DCT-domain priors.Source: §I · pp. 1–2
- IDCN directly estimates DCT-domain losses without DCT to build a color-to-color network for reducing color image compression artifacts.The authors frame it as a unified architecture for dual-domain correction.Source: §I · p. 2
- Its dual-domain correction unit uses an extractor–corrector framework, while dilated convolutions enlarge the receptive field to remove wide-range distortions.A pixel position labeling map is concatenated to the input so the network can use position information to refine the compressed image.Source: §I · p. 2
- IDCN-f is designed to handle a wide range of compression qualities with one model weight.Source: §I · p. 2
II. RELATED WORK
- Earlier dual-domain work used DCT-domain priors and improved gray-scale restoration, but the authors identify effective color restoration with DCT priors as an unresolved problem.The related-work discussion describes DDCN as a dual-domain convolutional network that improved performance on gray-scale images.Source: §I · p. 1, §II · p. 3
III. IMPLICIT DUAL-DOMAIN CONVOLUTION NETWORK
IDCN addresses the difficulty of using DCT-domain priors for color-image restoration by translating DCT-related losses into the pixel-feature domain, where corrections can be integrated with pixel-domain features.
III. IMPLICIT DUAL-DOMAIN CONVOLUTION NETWORK
- Conventional DCT methods directly recover coefficients, a difficult fit for color images because luminance and chrominance use different compression operations.Chrominance is downsampled before DCT, and its quantization table differs from luminance's.Source: A. Implicit Translation from… · p. 3
- A 2D convolution translates DCT-domain loss into pixel-domain loss in YCbCr, avoiding the element-wise operation and reducing calculations by approximately 40%.The translation uses the quantization table and estimated relative quantization loss.Source: A. Implicit Translation from… · pp. 3–4
- The estimated luminance and chrominance losses are translated back into the pixel-feature domain to preserve the nonlinear feature-processing path and support skip connections.The passage also notes that each estimated loss vector represents an 8 × 8 patch.Source: A. Implicit Translation from… · p. 4
- The authors call this pixel-feature-to-pixel-feature operation using DCT-domain priors implicit dual-domain translation.Source: A. Implicit Translation from… · p. 4
B. Position Labeling
Position labeling supplies pixel-location information reflecting the periodic variation of compression deviations; the authors simplify the map for single-quality training because unchanged information can mislead the network.
B. Position Labeling
- Deviation standard deviations vary by position within DCT patches, with corner values larger than middle values and an 8 × 8 cycle in 16-pixel patches.The distribution regularity was reported across tested datasets, though standard-deviation ranges differed.Source: B. Position Labeling · p. 4
- The labeling map assigns standard-deviation values by pixel position and has the same spatial size as the input image.For color images, the full map has three channels and a minimum period of 16 pixels per axis.Source: B. Position Labeling · p. 4
- For single-quality networks, the authors simplify the multi-channel map because unchanged labeling information could mislead training toward map features rather than image features.They used the simplified map for single-quality training and a multi-channel map for flexible-network training.Source: B. Position Labeling · pp. 4–5
C. Architecture
IDCN processes the input through feature encoding, a correction baseline made of DCUs, and feature decoding to produce the output image.
C. Architecture
- The feature encoder maps the input into pixel-feature representations using two convolution layers, each linked to a ReLU layer.The first convolution uses a 5 × 5 kernel; the second uses a 3 × 3 kernel.Source: C. Architecture · p. 5
- A series of DCUs forms the correction baseline, with each unit applying a dual-domain correction operation to the preceding feature representation.Source: C. Architecture · p. 5
- The feature decoder maps corrected features to the final color image through two convolution layers, with a ReLU after the first.The first decoder convolution uses a 3 × 3 kernel and the second a 5 × 5 kernel.Source: C. Architecture · p. 5
D. Dual-domain Correction Unit
The dual-domain correction unit builds on an extractor–corrector residual design, sharing a dense feature extractor across DCT-domain and pixel-domain correction branches.
D. Dual-domain Correction Unit
- The DCU adds DCT-domain and pixel-domain corrections to its input feature through a residual connection.Both correction branches use the shared feature extractor.Source: D. Dual-domain Correction Unit · p. 6
- The shared extractor uses densely connected convolutions and a transition block to compress high-dimensional extracted features.Dilated convolution layers enlarge the extractor's receptive field.Source: D. Dual-domain Correction Unit · p. 6
- The DCT branch estimates relative quantization loss for luminance and integrated chrominance, then translates those estimates into pixel-feature corrections.It uses separate estimators because the chrominance channels share a quantization table; the chrominance estimator uses dilation rate 2.Source: D. Dual-domain Correction Unit · p. 6
- The pixel-domain branch uses a convolutional corrector, while a scaling layer scales both branch corrections at 0.1 to avoid gradient explosion.Source: D. Dual-domain Correction Unit · p. 6
A. Implementation Details
IDCN uses a repeated DCU network configuration and is evaluated on standard image datasets with full-color quality metrics. Training uses MSE and Adam with a staged learning-rate schedule.
A. Implementation Details
- IDCN uses 8 DCUs with K = 64 and L = 8; most convolution layers have 64 filters.The REU convolution layers also use 64 filters to output relative quantization losses for 64 DCT coefficients.Source: A. Implementation Details · p. 6
- Figure 8 compares estimated DCT-domain losses with true DCT-domain losses.The comparison includes luminance and combined chrominance estimates and true luminance, Cb, and Cr losses.Source: A. Implementation Details · p. 7
- Testing uses 200 BSDS500 images, 29 LIVE1 images, and 143 WIN143 images, after training on 900 DIV2K images.The DIV2K test images are excluded because their ground truth is unavailable.Source: A. Implementation Details · p. 7
- Evaluation measures full-color PSNR and the mean RGB-channel SSIM and PSNR-B.The metrics are computed with standard MATLAB library functions.Source: A. Implementation Details · p. 7
- Training minimizes MSE with Adam and reduces the learning rate when validation loss plateaus.Training stops after five consecutive epochs with less than 0.001 dB validation-loss decrease at the minimum learning rate.Source: A. Implementation Details · p. 7
B. Ablation Investigation
The ablation examines position labeling, correction-channel choices, and dilated convolution. Simplified labeling and simple correction improve performance, while dilation offers little measured enhancement but helps remove wide-range artifacts.
B. Ablation Investigation
- The ablation varies position labeling, multi-channel correction, and dilated convolution, including simplified or multi-channel labels and full or simple correction.Full correction uses Y, Cb, and Cr; simple correction uses luminance and combined chrominance.Source: B. Ablation Investigation · p. 7
- Table I reports the PL and MCC ablation on LIVE1 at q = 20 using PSNR, with four networks sharing the same settings.Figure 10 presents subjective evaluations at q = 10 for three listed images.Source: B. Ablation Investigation · p. 7, Figure 10 · p. 8
- Simple correction improves performance by simplifying the network, while full correction may perform worse because additional DCT correction branches can scatter backpropagated loss.Simplified labeling also improves performance by adding pixel-position information without misleading training.Source: B. Ablation Investigation · pp. 7, 9
- Dilated convolution provides little performance enhancement, but enlarging the extractor receptive field helps remove wide-range artifacts.Figure 9 shows correction maps from dual-domain, DCT-domain, pixel-domain, luminance, and combined chrominance outputs.Source: Figure 9 · p. 7, B. Ablation Investigation · p. 9
C. Investigation of Dual-domain Learning
Among pixel-only, luminance-DCT, and dual-domain learning, IDCN-D performs best across PSNR, SSIM, and PSNR-B. Its DCT correction maps are smoother, and estimated DCT losses concentrate in high-frequency areas.
C. Investigation of Dual-domain Learning
- IDCN-D achieves the best PSNR, SSIM, and PSNR-B results, while IDCN-Y is slightly better than IDCN-P.Table II identifies the comparison as an investigation of dual-domain learning.Source: C. Investigation of Dual-domain Learning · p. 9
- DCT-domain correction maps are smoother than pixel-domain maps, where some high-frequency artifacts remain.Luminance and combined-chrominance correction maps are similar, differing mainly around edges.Source: C. Investigation of Dual-domain Learning · p. 9
- The luminance and combined-chrominance correction maps are similar because their structures and training gradients are similar.Dilated convolution for combined chrominance loss estimation matches compression downsampling and probably contributes to low-frequency output.Source: C. Investigation of Dual-domain Learning · p. 9
- Estimated DCT coefficient losses focus on high-frequency areas, matching the smoother outputs observed in the DCU analysis.Their mean norms are statistically close to true losses around high frequencies; estimated loss norms are around zero.Source: C. Investigation of Dual-domain Learning · pp. 9–10
D. Comparison with State-of-the-Art Methods
IDCN improves on compared methods across tested color-image datasets, compression qualities, and metrics, with a larger gap at lower q. It also achieves state-of-the-art results on luminance-channel comparisons, while running at a speed similar to simplified RDN.
D. Comparison with State-of-the-Art Methods
- Qualitatively, IDCN produces more defined edges and accurate colors and is more robust to wide-range color distortion.Other methods retain block effects, blurring, color aberrations, or color banding; IDCN results are described as more visually pleasing.Source: D. Comparison with State-of-the-Art… · p. 10
- On a 2080Ti GPU, IDCN runs at a speed similar to simplified RDN and is slightly slower than MemNet and S-Net.Downsampling gives MWCNN and DWP-SDNet excellent computational efficiency.Source: Figure 11 · p. 10, D. Comparison with State-of-the-Art… · p. 11
E. Robust Compression Artifact Reduction
IDCN-f uses endpoint labeling maps to handle a range of compression qualities, with high performance across that range and a slight sacrifice versus quality-specific IDCN models. The conclusion also reports IDCN’s superiority over state-of-the-art methods and highlights IDCN-f’s potential for practical compression artifact reduction.
E. Robust Compression Artifact Reduction
- IDCN-f performs at a high level across compression qualities [5,20], with slight degradation versus quality-specific models; its LIVE1 performance curves are shown in Fig. 12.The text also reports that IDCN-f fixes chromatic aberrations and texture details across a wide quality range.Source: Figure 12 · p. 11, E. Robust Compression Artifact Reduction · p. 11
- IDCN-f generates labeling maps between endpoint qualities by interpolating their maps, avoiding the need to calculate statistical priorities separately for every quality.The method is described for a quality range bounded by q_l and q_h.Source: E. Robust Compression Artifact Reduction · p. 11
V. CONCLUSIONS
- The conclusion reports IDCN’s superiority over state-of-the-art methods on widely used datasets and expanded WIN143, while IDCN-f handles variable qualities in subjective and objective evaluations.The proposed network uses an extractor-corrector framework and dual-domain correction; DCT-domain losses are estimated directly from extracted features without applying DCT.Source: §V · p. 12