{"id":"d4dcd86c-9dbd-4655-a62c-2ba570d6ed61","arxiv_id":"2506.21171","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A new DCT-domain enhancement framework exploits block-based and point-based correlations in JPEG coefficients to adapt pixel-domain networks, yielding 0.35 dB PSNR gain and 60.5% throughput gain.","lead":"This paper presents AJQE, a method that enhances JPEG images by working directly on the compressed DCT coefficients in the file, instead of decoding the image to pixels first. It reports 0.35 dB higher PSNR and 60.5% higher throughput than adapting the same neural networks in the pixel domain.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"PSNR gain may be driven by quantization-matrix side information that the pixel-domain baselines never receive; Tab. 4 never ablates this, so the 0.35 dB claim is unproven.","rationale":"I agree with the reader's weakest-assumption analysis: the missing ablation of QY/QC side information is the most load-bearing concern for the headline quality claim. The paper's own ablation study (Tab. 4) varies the internal modules but never removes the quantization-matrix input while keeping EMBC intact, so it cannot establish that the correlation modules, rather than the metadata, produce the 0.35 dB advantage. This is not a complaint about using JPEG metadata in a DCT-domain method; it is a complaint about the attribution of the gain to the proposed modules and the fairness of the comparison to pixel-domain baselines. The concrete test I propose directly isolates the side-information contribution by holding the architecture fixed and ablating only the Q-matrix input. Since the original verdict is already conditional on exactly this issue, my stress-test does not move the verdict.","tokens_in":47280,"tokens_out":8887,"duration_ms":106731,"concrete_test":"Retrain the AJQE-w/FBCNN model on the Sec. 5.1 setup with QY/QC replaced by a constant all-ones matrix in EMBC's kernel-generation path (all other components and training settings unchanged), and compare BSDS500 PSNR at QF 10/20/30/40 against Table 2. If the average PSNR deficit relative to the original AJQE-w/FBCNN is large (e.g., >0.2 dB), the reported gain is substantially side-information-driven; if the deficit is negligible, the correlation modules carry the claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quality claim (Abstract; Sec. 5.2) attributes an average 0.35 dB PSNR gain to the proposed EMBC/EMPC modules, but the comparison does not control for the fact that EMBC's block-based convolution consumes the JPEG quantization matrices QY and QC as inputs (Sec. 4.1, Fig. 6). The retrained pixel-domain baselines (AR-CNN, DCAD, DnCNN, FBCNN, ARCRL) do not receive this metadata; FBCNN must predict QF, while QGAC (a DCT-domain baseline) explicitly uses quantization information. Table 4 ablates EMBC/EMPC to ResBlocks and replaces block convolution with vanilla convolution, but no arm removes QY/QC from EMBC while holding the rest of the method fixed, so the side-information contribution cannot be separated from the proposed correlation mechanism. In addition, Tables 2 and 3 report only the AJQE variants with parenthetical deltas; the actual pixel-domain baseline numbers are absent, and Fig. 1's illustrative FBCNN baseline (26.13 dB, AJQE 26.81 dB on BSDS500 QF10) does not reconcile with Table 2's implied baseline for FBCNN (27.69 dB - 0.29 dB = 27.40 dB). The direction of the concern is therefore not hypothetical: the headline gain is not auditable from the printed tables.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes AJQE, a method for JPEG quality enhancement that operates directly on DCT coefficients. The authors identify two types of latent correlations in DCT data—block-based and point-based correlations—and design two modules, EMBC and EMPC, to exploit them. These modules allow existing pixel-domain enhancement networks (AR-CNN, DCAD, DnCNN, FBCNN, ARCRL) to be adapted to the DCT domain. The paper claims an average 0.35 dB PSNR improvement over the pixel-domain counterparts and a 60.5% increase in enhancement throughput, with experiments on several benchmarks and QF values.","tokens_in":47572,"tokens_out":4706,"duration_ms":51102,"significance":"If the claims hold, the paper offers a practical recipe for moving JPEG enhancement into the compressed domain, with substantial computational savings and a principled justification based on measured correlations. The correlation analysis is a useful empirical contribution, and the UMEP lemma is a clean identity with a proof included in the supplement. The authors retrain all baselines under consistent settings, which is more careful than typical in this area. However, the headline quality gain is not yet fully auditable because of a side-information confound and missing baseline numbers in the main tables.","major_comments":[{"comment":"The block-based convolution in EMBC generates its kernel and bias from the quantization matrices QY and QC (Sec. 4.1, Fig. 6). In the first stage, QY, Cb_DCT, and Cr_DCT are used; in the second stage, QC and Y_DCT are used. The pixel-domain baselines receive no such metadata: FBCNN is QF-blind and must predict QF, and the other pixel-domain models do not even condition on QF. The ablation study in Table 4 does not include an arm that removes QY/QC from EMBC while keeping the block-based convolution and the rest of the method fixed. The BConv-VConv arm replaces block-based convolution with vanilla convolution, which changes both the side information and the convolution structure, so it does not isolate the side-information effect. Therefore the average 0.35 dB improvement attributed to EMBC/EMPC in the Abstract and Sec. 5.2 may be partly or wholly due to explicit access to quantization step sizes, and the claim is not proven by the reported experiments.","section":"Sec. 4.1, Fig. 6, Table 4"},{"comment":"Tables 2 and 3 report only the AJQE variant numbers with parenthetical deltas (e.g., \"27.69 (+0.29)\"), but they do not print the actual pixel-domain baseline PSNR/SSIM values. The parenthetical improvements can therefore not be independently verified, and the average improvements quoted in Sec. 5.2 are not auditable from the tables. There is a direct inconsistency with Fig. 1: for BSDS500 QF=10, Table 2 lists AJQE w/ FBCNN as 27.69 dB and implies a pixel-domain FBCNN baseline of 27.69 - 0.29 = 27.40 dB, while Fig. 1 reports FBCNN as 26.13 dB and FBCNN w/ Ours as 26.81 dB. The JPEG baseline itself also differs (25.47 dB in Table 2 vs. 23.42 dB in Fig. 1). The authors should print the baseline numbers in the tables and reconcile the figure with the tables.","section":"Tables 2 and 3, Fig. 1"},{"comment":"The ablation study does not support the attribution of the quality gain to the proposed correlation modules as opposed to the extra side information. A controlled ablation is needed in which the EMBC block-based convolution is retained but QY/QC are removed from the kernel-generation process (e.g., generating the kernel from the DCT coefficients alone, or from a constant). Without such an arm, the 0.35 dB gain over pixel-domain baselines cannot be separated from the quantization-matrix advantage. Adding this experiment is essential to validate the central claim of the paper.","section":"Table 4"}],"minor_comments":[{"comment":"The EMPC compaction ratio α is a free parameter, but its value is never specified in the paper. The authors should state the α used in the main experiments and whether the learnable or non-learnable compaction strategy was employed.","section":"Sec. 4.2, Sec. 5.1"},{"comment":"The table captions say \"Performance improvements over the pixel-domain model are shown in parentheses,\" but the columns are labeled with the AJQE variant, not the corresponding pixel-domain baseline. Each column should explicitly name the baseline model it is adapted from.","section":"Tables 2 and 3 captions"},{"comment":"The caption states that the comparison includes JDEC and other methods, but the figure as printed appears to show only FBCNN and FBCNN w/ Ours. Confirm that all baselines are actually plotted or adjust the caption.","section":"Fig. 1 caption"},{"comment":"For reproducibility, the authors should provide the kernel sizes and layer counts for the block-based convolution in EMBC and the exact architecture of the kernel-generation network, since these details are only sketched in Fig. 6.","section":"Sec. 5.1"},{"comment":"The ground-truth DCT coefficients are obtained by compressing with QF=100. Please clarify how images whose dimensions are not multiples of 8 are handled in the DCT-domain supervision, since padding or cropping can affect the DCT-domain loss.","section":"Eq. (3)"}],"recommendation":"major_revision","confidential_remarks":"The paper is well-structured and the empirical correlation analysis is interesting. The main issue is the unablated side-information advantage: the DCT-domain models get the quantization matrix while pixel-domain baselines do not. This is a correctable but load-bearing flaw. Also, the missing baseline numbers in Tables 2/3 and the inconsistency with Fig. 1 will make it hard for readers to verify the headline numbers. I would recommend asking for a controlled ablation and a full reporting of baseline values."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick read on arXiv:2506.21171. The AJQE paper adapts pixel-domain JPEG enhancement models to work directly on DCT coefficients, using two modules that exploit block-based and point-based correlations. That framing is genuinely new, and the empirical analysis in Sec. 3 (Moran's I / Geary's C on coefficient maps) is a solid motivation. The UMEP lemma (DCT-domain upsampling equivalent to pixel-domain NN upsampling) is a neat, correct piece of math. The efficiency story is real: skipping the pixel decode gives a 60.5% throughput gain and fewer FLOPs, which matters for web-scale JPEG processing.\n\nThe soft spot is the headline quality claim. EMBC consumes the quantization matrices QY and QC as inputs to generate its kernels. The pixel-domain baselines never get that metadata; FBCNN has to predict QF. So part of the 0.35 dB gain could simply be from having the quantization table handed to the model. Table 4 ablates EMBC to ResBlocks and swaps block convolution for vanilla, but no arm removes QY/QC while holding the rest fixed. That means the side-information effect is not separated from the correlation effect. Also, Tables 2 and 3 only report deltas, not the actual baseline numbers, which makes the comparison hard to verify from the printed tables. The Fig. 1 FBCNN baseline (26.13 dB) doesn't match the implied baseline from Table 2 (27.40 dB). That looks like an inconsistency that needs cleaning up.\n\nI don't think this is a fatal flaw. The method is still a reasonable engineering contribution, and the efficiency gain is independent of the side-information question. But the 0.35 dB PSNR claim, as stated, is not yet supported. The authors should add an ablation that feeds QY/QC to the pixel-domain baselines (or removes them from EMBC) and report full baseline numbers.\n\nWould I cite it? Only if the side-info ablation comes out, or for the UMEP lemma. It deserves a serious referee: the idea is strong enough, and the proof is easy to check. My recommendation: send it to review, but flag the side-information confound as a required revision.","headline":"A useful DCT-domain adaptation framework with a clean efficiency win, but the 0.35 dB quality claim is confounded by the quantization-matrix side information.","tokens_in":48073,"tokens_out":2718,"would_cite":false,"duration_ms":30735,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"JPEG artifact removal moves into DCT domain and beats pixel-domain rivals.","keywords":["JPEG quality enhancement","DCT-domain enhancement","block-based correlation","point-based correlation","coefficient maps","quantization matrix","compression artifact removal","DCT-domain upsampling"],"falsifier":"Retrain the DCT model with EMBC's kernel generator receiving only the DCT coefficients and not $Q_Y$/$Q_C$ (or train a pixel-domain baseline that is given the true quantization matrix), and compare PSNR; if the 0.35 dB gap disappears, the side information rather than the block- and point-based correlations is the cause.","tokens_in":47098,"feed_emoji":"🖼️","tokens_out":5296,"duration_ms":54695,"temperature":0.7,"pith_summary":"JPEG compresses images by quantizing DCT coefficients, and that quantization is the source of blocking and ringing artifacts. The paper argues that the standard way to fix these artifacts—decode to pixels and enhance with a convolutional network—is wasteful and misses structure that lives in the compressed domain itself. It claims that DCT coefficients are not weakly correlated as previously assumed: they carry two latent correlations, a block-based one across neighboring $8\\times8$ blocks and a point-based one within coefficient maps that align the same frequency across all blocks. Exploiting both lets the authors adapt existing pixel-domain enhancement models to work directly on the entropy-decoded coefficient blocks. The result, AJQE, is claimed to improve PSNR by an average of 0.35 dB over the pixel-domain counterparts and to raise enhancement throughput by 60.5% by skipping inverse DCT and color conversion.","feed_headline":"JPEG artifact removal moves into DCT domain and beats pixel-domain rivals.","feed_subtitle":"Exploiting two latent correlations in DCT coefficients skips full decoding, gains 0.35 dB and 60.5% faster throughput.","key_machinery":"The load-bearing object is the reorganized DCT tensor. In EMBC, a learnable block-based convolution—a convolution whose kernel and bias are generated from the dequantization matrix $Q_Y$ (or $Q_C$) and the other color channels—turns each dequantized $8\\times8$ block into a block feature, passes it through a pixel-domain enhancement backbone, then maps back to DCT coefficients, thereby exposing block-based correlation to spatially invariant convolutions. In EMPC, the DCT tensor is remapped into 64 coefficient maps (one per frequency), compacted to the low-frequency (or a learnably selected) subset, enhanced by a pixel-domain model, and remapped back, exposing point-based correlation. A third piece, UMEP, supplies a fixed DCT-domain upsampling matrix $\\hat U_4 = T_8 U_4 T_4^\\top$ for chroma blocks that is algebraically equivalent to pixel-space nearest-neighbor upsampling, avoiding the aliasing that naive coefficient upsampling produces.","core_discovery":"The paper's central claim is that JPEG artifact enhancement can be done wholly in the DCT domain, without fully decoding to pixels, at higher quality than the same models used on pixels. The enabling observation is empirical: raw DCT coefficients show almost no spatial autocorrelation (standard metrics are near the 'no correlation' levels), but two reorganized views of the same coefficients are strongly correlated. Dequantizing each $8\\times8$ block and computing weighted sums over the block yields a block-level summary whose neighboring values correlate strongly, while grouping the same frequency position across all blocks into 64 coefficient maps yields maps whose low-frequency entries are spatially smooth. On these two reorganizations the paper builds EMBC and EMPC, which wrap an off-the-shelf pixel-domain network between operations that map DCT blocks to and from these correlated representations and back. Across five pixel-domain models on three color and four grayscale test sets, the DCT-domain versions consistently outperform their pixel-domain versions, with an average of 0.35 dB PSNR gain (color) and 0.39 dB (grayscale), plus a 60.5% throughput increase and 45.9% fewer FLOPs, while also outperforming prior DCT-domain methods such as QGAC and JDEC.","pith_inferences":["If the gain is partly due to explicit access to the quantization matrix, the same wrapper idea could be applied to pixel-domain baselines by feeding them that same side information, which would isolate the correlation modules' true contribution; this ablation is not in the paper.","The coefficient-map view suggests a direct link to frequency-domain vision transformers: EMPC's compaction to the top $\\alpha\\%$ low-frequency maps is a hand-designed version of what a learned frequency attention might discover, so the module could be replaced by a learned selection with little loss.","Since block-based correlation emerges only after dequantization, the gain may shrink for very high quality factors near lossless, where quantized coefficients retain little of that latent structure; the paper's enhancement tables cover only QF 10 through 40.","The method outputs enhanced DCT coefficients, so it can pair naturally with transcoding pipelines that re-encode with a larger quantization matrix or with downstream JPEG-domain vision models that consume bitstreams directly."],"forward_implications":["Any existing pixel-domain JPEG enhancement network can be wrapped by EMBC/EMPC to operate in the DCT domain, yielding higher PSNR/SSIM and lower FLOPs than the same network in pixels.","Enhancement no longer requires the inverse DCT and YCbCr-to-RGB conversion for storage or transmission; the output can be entropy-coded directly to a smaller JPEG file.","The improvement holds across quality factors 10 through 40 and for both color and grayscale, so the method is not tuned to a single compression setting.","DCT-domain models derived this way also beat dedicated DCT-domain methods (QGAC, JDEC) by an average of 0.41 dB color and 0.48 dB grayscale PSNR.","Training cost drops by 38.0% and enhancement throughput rises 60.5% compared with the pixel-domain counterparts."],"supporting_citations":[{"why":"Supplies the DIV2K training dataset used to train all models and DCT-domain variants.","marker":"[1]"},{"why":"Defines the DCT basis used in the UMEP upsampling derivation and in the block transforms.","marker":"[2]"},{"why":"Supplies the BSDS500 test set used for the headline PSNR and throughput comparisons.","marker":"[4]"},{"why":"Provides AR-CNN, one of the pixel-domain models adapted to the DCT domain by the proposed modules.","marker":"[8]"},{"why":"Provides FBCNN, the main pixel-domain baseline and the backbone used in the ablation study.","marker":"[18]"},{"why":"Prior quantization-guided DCT-domain enhancement method that introduces the block-convolution concept and serves as a comparison baseline.","marker":"[10]"},{"why":"JDEC, the state-of-the-art DCT-domain baseline that the proposed method outperforms by 0.41 dB color and 0.48 dB grayscale.","marker":"[15]"},{"why":"Source of the coefficient-map extraction used to define point-based correlations and the EMPC module.","marker":"[23]"},{"why":"Supplies the sub-block conversion method that splits 8x8 chroma DCT blocks into 4x4 blocks for UMEP.","marker":"[28]"},{"why":"Another coefficient-map reference used to motivate point-based correlations in the DCT domain.","marker":"[42]"}],"fun_headline_variants":["DCT-domain JPEG enhancement beats pixel-domain, 60% faster","No full decode: DCT-domain JPEG gains 0.35 dB, 60% speed","Exploit DCT correlations: JPEG quality up 0.35 dB, 60% faster","Skip pixel decoding: DCT-domain JPEG enhancement betters rivals"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported quality gain is credited to the new correlation modules, but the DCT-domain models are also given the JPEG quantization matrix as explicit input while the pixel-domain baselines must estimate or infer it, and that side-information advantage is not ablated in the paper.","fun_headline_variants_meta":{"raw":{"variants":["DCT-domain JPEG enhancement beats pixel-domain, 60% faster","No full decode: DCT-domain JPEG gains 0.35 dB, 60% speed","Exploit DCT correlations: JPEG quality up 0.35 dB, 60% faster","Skip pixel decoding: DCT-domain JPEG enhancement betters rivals"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000784,"raw_usage":{"total_tokens":3480,"prompt_tokens":987,"completion_tokens":2493,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":603,"completion_tokens_details":{"reasoning_tokens":2407}},"tokens_in":603,"tokens_out":2493,"duration_ms":19755,"temperature":1.0,"reasoning_tokens":2407,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:32:09.284148+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the DCT model with EMBC's kernel generator receiving only the DCT coefficients and not $Q_Y$/$Q_C$ (or train a pixel-domain baseline that is given the true quantization matrix), and compare PSNR; if the 0.35 dB gap disappears, the side information rather than the block- and point-based correlations is the cause.","supporting_citations":[{"cited_title":"Ntire 2017 challenge on single image super-resolution: Dataset and study","cited_arxiv_id":null,"evidence_quote":"Supplies the DIV2K training dataset used to train all models and DCT-domain variants."},{"cited_title":"Ahmed, T","cited_arxiv_id":null,"evidence_quote":"Defines the DCT basis used in the UMEP upsampling derivation and in the block transforms."},{"cited_title":"Contour Detection and Hierarchical Image Segmentation.IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(5):898–916, 2011","cited_arxiv_id":null,"evidence_quote":"Supplies the BSDS500 test set used for the headline PSNR and throughput comparisons."},{"cited_title":"Compression Artifacts Reduction by a Deep Convolu- tional Network","cited_arxiv_id":null,"evidence_quote":"Provides AR-CNN, one of the pixel-domain models adapted to the DCT domain by the proposed modules."},{"cited_title":"Towards flex- ible blind jpeg artifacts removal","cited_arxiv_id":null,"evidence_quote":"Provides FBCNN, the main pixel-domain baseline and the backbone used in the ablation study."},{"cited_title":"Quantization Guided JPEG Artifact Correction","cited_arxiv_id":null,"evidence_quote":"Prior quantization-guided DCT-domain enhancement method that introduces the block-convolution concept and serves as a comparison baseline."},{"cited_title":"Jdec: Jpeg decoding via enhanced continuous cosine coefficients","cited_arxiv_id":null,"evidence_quote":"JDEC, the state-of-the-art DCT-domain baseline that the proposed method outperforms by 0.41 dB color and 0.48 dB grayscale."},{"cited_title":"JPEG Quantized Coefficient Recovery via DCT Domain Spatial-Frequential Transformer.IEEE Transactions on Image Processing, 33: 3385–3398, 2024","cited_arxiv_id":null,"evidence_quote":"Source of the coefficient-map extraction used to define point-based correlations and the EMPC module."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the sub-block conversion method that splits 8x8 chroma DCT blocks into 4x4 blocks for UMEP."},{"cited_title":"Learning in the frequency domain","cited_arxiv_id":null,"evidence_quote":"Another coefficient-map reference used to motivate point-based correlations in the DCT domain."}],"review_version":1}