{"id":"ad5effeb-b7f1-464a-9fd4-2196fca198f9","arxiv_id":"1909.05638","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A CNN can classify JPEG2000 images directly from their CDF 9/7 DWT coefficients, saving most of the decoding time while matching or slightly beating RGB-domain accuracy.","lead":"This paper trains image classifiers directly on the wavelet coefficients inside JPEG2000 compressed images, skipping the reconstruction step that most cloud vision systems perform. The method reports modest accuracy gains and large decoding-time savings on CIFAR-10 and Tiny ImageNet.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"DWT-domain augmentation equivalence (Eq. 3) is validated only on unquantized images; under JPEG2000 deadzone quantization, H'(dequantized coefficients) need not equal DWT of the spatially augmented compressed image, leaving the paper's key novelty unverified in its operating regime.","rationale":"The paper is a competent engineering contribution with clear speed benefits: Table 1 supports an over-80% reconstruction gain, and the model comparisons in Tables 3-4 show that DWT-domain inputs with the proposed augmentation achieve accuracy comparable to RGB inputs with fewer residual blocks. The reconstruction and speed claims are therefore not the soft spot. The reader's weakest assumption correctly identifies the augmentation equivalence as the load-bearing point: Eq. (3) and Figure 3 establish the similarity transform only for unquantized, exactly invertible DWT coefficients, while the actual training and inference data pass through the JPEG2000 deadzone quantizer. Because the quantization operation is nonlinear and does not commute with the augmentation operator, the augmented training samples are not exact simulations of compressed spatial augmentations. The paper provides no measurement of this mismatch, so the theoretical basis for the main novel component remains unverified in the lossy regime. Importantly, this concern does not refute the empirical accuracy results; the augmentation could still be effective as an approximate transform-domain augmentation. It does mean the central claim should be accepted only conditionally, pending either a demonstration that the quantization mismatch is small or a restatement of the augmentation as approximate. Hence the verdict remains CONDITIONAL, unchanged from the reader's assessment.","tokens_in":8072,"tokens_out":6384,"duration_ms":77278,"concrete_test":"Take a sample of CIFAR-10 images and JPEG2000-encode them at the compression ratios used in the paper (r = 0, 5, 10, 15). For each image, compute two DWT-domain inputs: (i) H' applied to the dequantized coefficients extracted from the compressed image, and (ii) the coefficients extracted from a JPEG2000 encode of the spatially augmented RGB image H(X). Measure the normalized L2 error between (i) and (ii), and compare the predictions of a trained DWT-domain classifier on both sets. If the classifier disagreement is negligible, the quantization mismatch is benign; if not, Eq. (3)'s exactness claim fails in the operating regime and the paper must either establish that the mismatch is harmless or reframe the augmentation as approximate rather than exact.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central novelty is DWT-domain augmentation via the similarity transform H' = A^{-1}HA in Eq. (3), which is claimed to make spatial augmentations applicable directly to JPEG2000 coefficients. The derivation assumes that the DWT is exactly invertible and that XwA = A^T X A holds, and Figure 3 demonstrates 'exactly same' behavior only for an unquantized image. In the actual JPEG2000 pipeline, however, the classifier inputs are dequantized coefficients, not exact DWT coefficients: quantization with a deadzone is nonlinear, so applying H' to dequantized coefficients generally does not produce the coefficients that would result from compressing the spatially augmented RGB image. The paper does not quantify this mismatch or test its effect on training. This is load-bearing because the proposed augmentation is the main source of accuracy improvement over plain DWT inputs (about 2.5% on CIFAR-10 and 1.5% on Tiny ImageNet) and is the reason DWT-domain classification can match or slightly beat RGB-domain accuracy. The end-to-end accuracy numbers may still be valid—the augmentation could work as an approximate DWT-domain regularizer—but the stated theoretical justification and the claim that the augmentation exactly reproduces spatial transforms are unestablished under lossy JPEG2000 compression.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses cloud-based image classification of JPEG2000-compressed images. The authors propose bypassing the RGB reconstruction step by feeding the CDF 9/7 DWT coefficients, extracted from inside the JPEG2000 decoder after dequantization, directly to a CNN classifier. They introduce a DWT-domain augmentation rule based on a similarity transform H' = A^{-1}HA (Eq. 3), claim that conventional spatial augmentations are ineffective in the DWT domain, and demonstrate on CIFAR-10 and Tiny ImageNet that shallower DWT-domain models can match or slightly exceed RGB-domain accuracy while saving more than 80% of decoding time. They also study performance under varying JPEG2000 compression ratios and show that fine-tuning from a pretrained model saves training time.","tokens_in":8328,"tokens_out":3618,"duration_ms":38003,"significance":"If the central claims hold, the work is practically valuable: skipping inverse DWT and color conversion gives a large, image-size-dependent decoding speedup, and the proposed augmentation would be a useful tool for training classifiers directly on JPEG2000 coefficients. The paper's strengths are its clear system motivation, the explicit reconstruction-gain measurement (Table 1), and the fact that the DWT inputs are obtained from a real codec (OpenJPEG) rather than from synthetic clean coefficients. The main theoretical contribution, Eq. (3), is clean for an exactly invertible linear transform, but its validity in the actual lossy JPEG2000 operating regime is not established, which is the key risk to the claimed accuracy improvements. The empirical speed benefits are credible; the accuracy improvements over RGB are small and not always statistically separated.","major_comments":[{"comment":"The proposed augmentation H' = A^{-1}HA assumes that the DWT is an exactly invertible linear operator and that the network input XwA is the exact DWT of the spatial image Xs. In the actual JPEG2000 pipeline, however, the classifier inputs are dequantized coefficients; the deadzone quantizer is nonlinear and non-invertible. Applying H' to dequantized coefficients will generally not produce the same coefficients that would be obtained by compressing the spatially augmented RGB image at the same rate. The paper validates the 'exactly same' claim only on an unquantized image (Figure 3) and provides no analysis or experiment quantifying the mismatch under lossy compression (r > 0). Since the proposed augmentation is the main source of the accuracy gain over plain DWT inputs (about 2.5% on CIFAR-10 and 1.5% on Tiny ImageNet), the central contribution is unverified in its operating regime. The authors should either prove that the quantization error is small under the augmentation, or empirically compare the proposed augmented training against training with exactly matched compressed-augmented samples across compression ratios.","section":"Section 3.3, Eq. (3), Figure 3"},{"comment":"The claim that the proposed method achieves 'better accuracy' than RGB-domain classification is not statistically well supported. On CIFAR-10 the difference is 91.92% vs 91.70% with standard deviations of ±0.11 and ±0.07, and on Tiny ImageNet top-1 it is 67.56% vs 67.35% with ±0.09 and ±0.11. In both cases the intervals overlap and the differences are on the order of 1-2 combined standard deviations. The robust and consistent improvements are between CDF 9/7Aug and plain CDF 9/7, not between DWT and RGB. The abstract and conclusion should be revised to state the comparison more precisely, or the authors should provide a significance test or more runs.","section":"Section 4.2, Tables 3 and 4"},{"comment":"Several experimental details required for reproducibility are missing. The paper does not specify the exact JPEG2000 compression ratio r used for each accuracy table, the number of DWT decomposition levels, whether the level offset of 128 is applied before extracting coefficients, or how the 12 DWT subbands are scaled/normalized before being fed to the CNN. It is also unclear whether the reported training and inference speeds include the cost of the DWT extraction and the augmentation transform, or only the CNN forward/backward pass. These details are necessary to reproduce the claimed speedups and accuracy numbers, and should be stated explicitly.","section":"Section 4.2 and Table 1"}],"minor_comments":[{"comment":"The equation 'XsH = XsH' appears to be a typo; it should be written as a matrix product, e.g., XsH = Xs·H, for the subsequent derivation of Eq. (3) to be meaningful.","section":"Section 3.3, Eq. (2)"},{"comment":"There are several typographical errors: 'spatial RBG domain' should be 'RGB', and the Figure 3 caption contains 'fillped' instead of 'flipped'.","section":"Abstract and Figure 3 caption"},{"comment":"The caption states that the image 'Brownie' is of size 112×112, but no context is given for how this image relates to the CIFAR-10 or Tiny ImageNet datasets used in the experiments.","section":"Figure 3"},{"comment":"Reference [5] is a self-citation of the authors' prior work on DB1 wavelet coefficients; the paper should more clearly delineate what is new in the present work beyond changing the wavelet basis, especially because the augmentation method is the stated novelty.","section":"Section 2, reference [5]"},{"comment":"The code for the DWT matrix transformation is shared via a Google Drive link, which is not a stable archival repository; the authors should include the code as supplementary material or in a public repository with a persistent DOI.","section":"Section 3.7, footnote 2"}],"recommendation":"major_revision","confidential_remarks":"The main load-bearing issue is the quantization mismatch for the proposed augmentation, which is the paper's core novelty. The end-to-end accuracy results may still be empirically valid, but the theoretical justification is incomplete. The accuracy advantage over RGB baselines is marginal, so the practical impact rests mostly on the decoding speedup, which is well supported. I would suggest the editor ask for an additional experiment on lossy JPEG2000 streams that either validates the augmentation against a ground-truth compressed-augmented baseline or quantifies the degradation as a function of compression ratio."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a genuinely useful paper. It's the first to show you can classify straight from CDF 9/7 coefficients harvested inside a standard JPEG2000 decoder, and the reported time savings from skipping reconstruction are real. The main soft spot is the augmentation theory: the similarity transform H' = A^{-1}HA is exact only if the DWT is lossless, but the actual inputs are deadzone-quantized and dequantized coefficients. Figure 3 only demonstrates exactness on an unquantized image. That's a genuine gap, but I don't think it sinks the paper. The augmentation clearly helps empirically (about 2.5 points on CIFAR-10, 1.5 on Tiny ImageNet), and even if it isn't precisely equivalent to spatial augmentation, it's evidently a good DWT-domain regularizer. The claim that it 'exactly reproduces' spatial transforms is overstrong; the authors should either condition it on no quantization or quantify the mismatch.\n\nWhat's actually new: taking CDF 9/7 coefficients from inside the JPEG2000 decoder and feeding them to a ResNet, plus the proposed augmentation transformations. Prior work used DCT or unquantized DWT; this is the first to work with the actual lossy-j2k coefficients. The reconstruction gain numbers (80%+ of decode time) are meaningful and grow with image size. Experiments are competently run on two datasets, with standard deviations over three runs, and the speed/accuracy trade-off curves are informative.\n\nSoft spots: the accuracy gains over RGB are tiny (0.2–0.3 pp), so the pitch should be speed and bandwidth, not accuracy. 'Shallower' is a bit loose: the DWT models use fewer conv layers but actually have more parameters (Table 2), though the smaller spatial resolution keeps them faster. Code: only the augmentation matrix code is linked; the full training pipeline isn't, so the central numbers aren't fully reproducible from the paper alone. The self-citation to Quannet is fine; it's directly relevant and not excessive.\n\nThe quantization-mismatch concern is real and load-bearing for the theoretical justification, but the paper's main claims — faster inference and comparable accuracy — hold up. The augmentation improves accuracy even if its exact-equivalence story is approximate. I'd send this to peer review and ask the authors to address the quantization effect, release more code, and tone down the exactness wording. For anyone working on compressed-domain vision, this is worth reading.","headline":"A useful compressed-domain classification paper that delivers real speed savings and a novel DWT augmentation, but the augmentation's exactness claim doesn't survive JPEG2000 quantization.","tokens_in":8845,"tokens_out":3615,"would_cite":true,"duration_ms":38880,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"JPEG2000 images classify directly from wavelet coefficients, skipping reconstruction.","keywords":["JPEG2000","CDF 9/7 wavelet","DWT-domain classification","compressed-domain inference","image augmentation","cloud image classification","bandwidth-limited networks","ResNet"],"falsifier":"Train a model on DWT coefficients produced by the similarity transform of heavily quantized JPEG2000 coefficients, and a second model on coefficients obtained by compressing spatially flipped or shifted originals at the same bitrate; if accuracy of the first falls behind the second as the quantization step grows, the exact-equivalence claim fails for lossy streams.","tokens_in":7876,"feed_emoji":"🖼️","tokens_out":7038,"duration_ms":62281,"temperature":0.7,"pith_summary":"This paper argues that a deep CNN can classify JPEG2000-compressed images directly from the CDF 9/7 wavelet coefficients already present in the compressed stream, so the inverse wavelet transform, color conversion, and level offset are unnecessary. On CIFAR-10 and Tiny ImageNet, it reports that this compressed-domain approach matches or slightly beats RGB-domain classification while using shallower networks, and that it removes more than 80 percent of decoder reconstruction work. The accuracy gain rests on a new augmentation recipe: rather than flipping or shifting DWT coefficients directly, the spatial flip or shift is conjugated by the wavelet matrix, preserving the geometric meaning of the operation. If the paper is right, JPEG2000 classification can be both faster and more accurate in networked, bandwidth-limited settings without changing the codec.","feed_headline":"JPEG2000 images classify directly from wavelet coefficients","feed_subtitle":"Skipping the inverse wavelet transform saves over 80 percent of decoder time while matching or beating RGB accuracy.","key_machinery":"The key object is the matrix $A$ expressing the level-1 CDF 9/7 DWT as a separable linear transform, factored through the codec's lifting steps as $A = P_1 U_1 P_2 U_2 S$ (predictions, updates, and de-interleaving). Because $A$ is invertible, a spatial augmentation $H$ is replaced in the DWT domain by the conjugated transform $H' = A^{-1} H A$, which is what makes flip and shift augmentation meaningful for compressed-domain inputs. The same matrix representation also lets the authors compute DWT coefficients quickly during training, avoiding a slow wavelet pass over each image.","core_discovery":"The central claim is that dequantized CDF 9/7 DWT coefficients taken from inside a JPEG2000 decoder are a sufficient, and with the right augmentation a better, input representation for CNN classification than reconstructed RGB images. Representing the level-1 DWT as an invertible matrix $A$, the authors define augmentation by the similarity transform $H' = A^{-1} H A$, so that a horizontal flip or shift of the DWT coefficients corresponds to the same geometric operation on the original image. On CIFAR-10 this yields 91.92 percent accuracy for a 21-convolution-layer DWT ResNet versus 91.70 percent for a 27-layer RGB ResNet, and Tiny ImageNet shows a smaller advantage; both use fewer layers in the DWT domain. The practical message is that the decoder can stop after dequantization and the harvested coefficients train models that generalize across compression ratios.","pith_inferences":["If the same coefficient-harvesting trick ports to other transform codecs, any decoder that exposes quantized coefficients could feed a classifier before reconstruction, thinning edge-to-cloud inference pipelines.","The augmentation identity assumes the encoder's DWT is exactly the linear operator $A$; under heavy JPEG2000 quantization the augmented coefficients will not exactly equal coefficients of a spatially augmented original, so a quantization-aware version of $H'$ is the natural next test.","Because the inputs are level-1 subbands, the same pipeline could support progressive or multi-resolution classification, using only the low-frequency subband for a quick decision and adding detail subbands when bandwidth allows; this goes beyond what the paper tests."],"forward_implications":["A JPEG2000 decoder can hand the dequantized DWT coefficients directly to a classifier, eliminating the inverse DWT and color conversion; the reported reconstruction saving is over 80 percent of decode time and grows with image size.","Shallower ResNets suffice in the DWT domain, with CIFAR-10 using 21 convolution layers instead of 27 and Tiny ImageNet using 31 instead of 40, matching or improving accuracy while speeding up training and inference.","The proposed DWT-domain augmentation recovers more than two percentage points of top-1 accuracy on CIFAR-10 compared with naive augmentation of DWT coefficients, making compressed-domain training viable.","A model trained on uncompressed coefficients can be fine-tuned for lower-bandwidth channels, cutting training time by about 75 percent while improving accuracy, and the DWT-domain accuracy advantage persists as bandwidth shrinks."],"supporting_citations":[{"why":"Establishes the precedent of classifying JPEG/DCT compressed images without reconstruction, which this paper extends to JPEG2000 DWT.","marker":"[1]"},{"why":"Supplies the DB1-wavelet bandwidth-limited classification baseline and the unquantized CDF 9/7 accuracy figure the authors aim to beat.","marker":"[5]"},{"why":"Gives the separable DWT matrix representation underlying Eq. (1) and the proposed similarity-transform augmentation.","marker":"[16]"},{"why":"Provides the ResNet architecture and bottleneck design used for all RGB and DWT-domain classifiers.","marker":"[15]"},{"why":"Documents JPEG2000's irreversible CDF 9/7 compression path and decoder structure that the coefficient harvesting exploits.","marker":"[13]"},{"why":"Describes EBCOT block coding, the JPEG2000 entropy stage whose output the decoder must process before coefficients are available.","marker":"[14]"},{"why":"Supplies ImageNet and its Tiny ImageNet subset as the larger benchmark for the reported accuracy comparisons.","marker":"[7]"},{"why":"Supplies CIFAR-10, the primary dataset for the speed-versus-accuracy comparisons and augmentation study.","marker":"[8]"}],"fun_headline_variants":["Classify JPEG2000 images without decoding pixels","Direct DWT domain CNN classification for JPEG2000","Faster JPEG2000 classification using wavelet coefficients","Skip JPEG2000 reconstruction: classify from DWT directly","JPEG2000 CNN: no inverse wavelet transform needed"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the codec's wavelet transform is a perfect linear operator captured by the matrix $A$, so that flipping or shifting DWT coefficients via $A^{-1} H A$ produces exactly the coefficients of a flipped or shifted original image; JPEG2000's lossy quantization breaks that exact equivalence.","fun_headline_variants_meta":{"raw":{"variants":["Classify JPEG2000 images without decoding pixels","Direct DWT domain CNN classification for JPEG2000","Faster JPEG2000 classification using wavelet coefficients","Skip JPEG2000 reconstruction: classify from DWT directly","JPEG2000 CNN: no inverse wavelet transform needed"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000204,"raw_usage":{"total_tokens":1376,"prompt_tokens":915,"completion_tokens":461,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":531,"completion_tokens_details":{"reasoning_tokens":385}},"tokens_in":531,"tokens_out":461,"duration_ms":4912,"temperature":1.0,"reasoning_tokens":385,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T05:00:50.376192+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a model on DWT coefficients produced by the similarity transform of heavily quantized JPEG2000 coefficients, and a second model on coefficients obtained by compressing spatially flipped or shifted originals at the same bitrate; if accuracy of the first falls behind the second as the quantization step grows, the exact-equivalence claim fails for lossy streams.","supporting_citations":[{"cited_title":"Faster neural networks straight from jpeg,","cited_arxiv_id":null,"evidence_quote":"Establishes the precedent of classifying JPEG/DCT compressed images without reconstruction, which this paper extends to JPEG2000 DWT."},{"cited_title":"Quannet: Joint image compression and classification over the channels with limited bandwidth,","cited_arxiv_id":null,"evidence_quote":"Supplies the DB1-wavelet bandwidth-limited classification baseline and the unquantized CDF 9/7 accuracy figure the authors aim to beat."},{"cited_title":"Digital image processing [m],","cited_arxiv_id":null,"evidence_quote":"Gives the separable DWT matrix representation underlying Eq. (1) and the proposed similarity-transform augmentation."},{"cited_title":"Deep residual learning for image recognition,","cited_arxiv_id":null,"evidence_quote":"Provides the ResNet architecture and bottleneck design used for all RGB and DWT-domain classifiers."},{"cited_title":"Taubman and M","cited_arxiv_id":null,"evidence_quote":"Documents JPEG2000's irreversible CDF 9/7 compression path and decoder structure that the coefficient harvesting exploits."},{"cited_title":"High performance scalable image compression with ebcot,","cited_arxiv_id":null,"evidence_quote":"Describes EBCOT block coding, the JPEG2000 entropy stage whose output the decoder must process before coefficients are available."},{"cited_title":"ImageNet Large Scale Visual Recognition Challenge,","cited_arxiv_id":null,"evidence_quote":"Supplies ImageNet and its Tiny ImageNet subset as the larger benchmark for the reported accuracy comparisons."},{"cited_title":"Imagenet classification with deep convolutional neural networks,","cited_arxiv_id":null,"evidence_quote":"Supplies CIFAR-10, the primary dataset for the speed-versus-accuracy comparisons and augmentation study."}],"review_version":1}