{"id":"7b96d761-03c0-4311-9079-279dbe4db8a5","arxiv_id":"1908.09288","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"LLISE is a Locally Linear Embedding variant that uses an SSIM-based distance instead of squared error and is reported to separate image distortion types better than LLE.","lead":"The authors replace the squared Euclidean distance in Locally Linear Embedding with a structural similarity distance, yielding a block-wise 'image structure manifold' for recognizing image distortions. The idea could benefit image quality assessment, but the experiments are small and the comparison to LLE is uneven.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline comparison is confounded: LLISE is evaluated block-wise with majority voting while LLE/kernel LLE are applied to whole images, so the reported advantage may stem from evaluation protocol rather than SSIM distance.","rationale":"The reader's weakest assumption about the zero-mean SSIM formula in the embedding objective is technically sound and also relevant: Eqs. (12)-(14) apply Eq. (2) to embedded points y_j,i that are not centered over their p coordinates, so the embedding loss is a normalized squared distance rather than a perceptual SSIM distance. However, that issue primarily weakens the theoretical interpretation of the method, not necessarily its empirical performance if the loss is redefined as a normalized quadratic. The more decisive problem for the central claim is the experimental confound: LLISE is evaluated block-wise with majority voting while the LLE baselines are described as whole-image methods. A superiority claim cannot be attributed to the SSIM distance unless the baseline is given the same block-level protocol. The concrete test would settle this by running block-wise LLE and kernel LLE. The verdict remains conditional: the paper should be accepted only after this fair comparison is performed and reported, alongside the zero-mean SSIM issue and the SVD projection typo noted by the reader.","tokens_in":19403,"tokens_out":6867,"duration_ms":73914,"concrete_test":"Re-run the comparison with LLE and kernel LLE on exactly the same block-wise protocol used for LLISE: for each block index i, compute a k-NN graph across the 121 training images using Euclidean distance on the centered 8x8 blocks, solve the standard LLE reconstruction and embedding for that block (k=10, p=4), then classify each block with 1NN in the embedded space and take a majority vote per image. If block-wise LLE or kernel LLE matches or exceeds LLISE/kernel LLISE on the confusion matrices and the out-of-sample Table 1, then the reported advantage is an artifact of evaluation granularity. If LLISE still outperforms block-wise LLE, the central claim survives this concern.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is empirical: 'Except for impulse noise, LLISE and kernel LLISE had better performance compared to LLE and kernel LLE' (Section 4). The evidence for this claim is not controlled. LLISE is explicitly block-wise: images are partitioned into 8x8 blocks, each block is embedded, each block is classified by 1NN, and the image label is decided by majority vote over 4096 blocks. In contrast, the paper states in the out-of-sample discussion that 'LLE does not perform block-wise and thus it has only one recognition label for the whole image.' If LLE and kernel LLE were also applied to whole 512x512 images in the training-phase confusion matrices of Fig. 2, then the comparison conflates two factors: (1) the proposed SSIM-based distance in the reconstruction and embedding objectives, and (2) the information advantage of block-level patch statistics plus majority voting. A whole-image LLE embedding of 121 images in R^262144 is also numerically fragile and likely to behave poorly for reasons unrelated to the distance metric. Therefore, the headline claim does not isolate the contribution of replacing the l2 norm with SSIM distance. This is a load-bearing concern because the paper's novelty and promised 'bridge between manifold learning and image fidelity assessment' rest on that empirical superiority claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Locally Linear Image Structural Embedding (LLISE) and its kernel variant as alternatives to Locally Linear Embedding (LLE) for learning what the authors call an 'image structure manifold.' The method replaces the squared Euclidean distance in LLE with an SSIM-based distance for both the reconstruction of each image block from its k nearest neighbors and the embedding step, with the weights computed by ADMM and the embedding constrained to have zero mean and unit covariance. The authors also provide an out-of-sample embedding procedure and a kernelized version. The experiments use 121 distorted versions of a single Lena image (six distortion types plus the original) and 12 out-of-sample images; distortion type is recognized by a 1NN classifier in the learned embedding, with majority voting over blocks for LLISE and one label per whole image for LLE and kernel LLE.","tokens_in":19813,"tokens_out":5444,"duration_ms":60675,"significance":"If the method worked as claimed, replacing the l2 norm with a perceptually motivated SSIM distance in a manifold learning pipeline would be a useful contribution and would provide a concrete bridge between image quality assessment and manifold learning. The paper is clearly written and includes full derivations of the gradients and update rules plus supplementary material, which is valuable for reproducibility. However, the significance is currently limited by three issues: the embedding objective does not actually use an SSIM distance because the embedded vectors are not zero-mean over their coordinates; the projection step for the unit-covariance constraint contains a mathematical error (singular values should be sqrt(n), not n); and the central empirical claim rests on a single image and on a confounded comparison between block-wise LLISE and whole-image LLE. These problems are load-bearing because the paper's novelty and promised application depend on both the SSIM interpretation and the reported empirical advantage.","major_comments":[{"comment":"The SSIM distance in Eq. (2) is derived under the assumption that both vectors have zero mean over their coordinates, as stated in Section 1.1. In the embedding problem (12), however, the arguments Y_i^T 1_j and Y_i^T w_j,i are p-dimensional embedded vectors, and the zero-mean constraint in Eq. (12) is across the n images (sum_j y_j,i = 0), not across the p coordinates. Thus the objective in Eq. (14) is not an SSIM distance but a normalized squared distance with a constant c inherited from the block size q; it does not have the perceptual interpretation claimed. The authors need to either center each embedded vector across its p coordinates before applying Eq. (2), or define and justify a different distance for the embedding space.","section":"§2.1, Eqs. (2) and (14)"},{"comment":"The constraint (1/n) Y_i^T Y_i = I implies that the singular values of Y_i are sqrt(n), not n. The text derives this from the SVD of Y_i but then incorrectly concludes 'Sigma = nI'; with singular values set to n, one obtains (1/n)Y_i^T Y_i = nI, violating the stated constraint. The projection in Eq. (20) should set the singular values to sqrt(n). While this scaling error may not change 1NN classification if applied consistently, it is a concrete algebraic error in a central algorithmic step and should be corrected.","section":"§2.1, projection step following Eq. (18)"},{"comment":"The headline comparison is confounded. LLISE and kernel LLISE are evaluated block-wise with majority voting over all blocks, while Section 4 explicitly states that LLE and kernel LLE do not perform block-wise and provide one label for the whole image. The reported advantage over LLE therefore conflates the proposed SSIM-based distance with the information advantage of patch-level statistics and ensemble voting. A controlled comparison would apply both methods under the same block-wise or whole-image protocol, or at least include a block-wise LLE baseline. Without this control, the sentence 'Except for impulse noise, LLISE and kernel LLISE had better performance compared to LLE and kernel LLE' is not supported.","section":"§4, Fig. 2 and Table 1"},{"comment":"The experimental evidence is based on a single Lena image and 121 distorted versions of it, with no multiple source images, no random seeds or confidence intervals, and no statistical testing. The claim that the method is useful for discriminating image distortion types in general is therefore not established. The authors should evaluate on multiple natural images, report variability across images, and ideally compare against a block-wise LLE and other baselines under matched conditions.","section":"§4, training dataset"}],"minor_comments":[{"comment":"In Eq. (24), the notation r~wij appears to be a typo for r~wj,i; please make the subscript consistent with the rest of the paper.","section":"§3.1, Eq. (24)"},{"comment":"The replacement of the standard LLE constraint sum_r r~w = 1 with sum_r (r~w)^2 = 1 is motivated by numerical stability, but it changes the meaning of 'linear reconstruction' and removes translation invariance; a brief discussion of the effect of this change on the learned weights and on the out-of-sample reconstruction (22) would help the reader.","section":"§2.1, Eq. (3)"},{"comment":"The term 'image structure manifold' is introduced as a new concept but is never formally defined; it would be clearer to state exactly what geometric or topological property is claimed for the learned embedding.","section":"Abstract and Section 2"},{"comment":"For LLE and kernel LLE, the table gives only a single label without percentages, while LLISE rows give percentages; providing comparable quantitative information for the baselines would strengthen the comparison.","section":"§4, Table 1"},{"comment":"The out-of-sample embedding in Eq. (22) uses only the nearest-neighbor reconstruction weights from the training data; it would be helpful to state whether the embedding coordinates ry(t) are taken from the LLISE or kernel LLISE training embedding and how this choice affects kernel variants.","section":"§2.2, out-of-sample embedding"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to be a preprint of a conference paper (ICIAR). The core idea is interesting and the derivations are mostly transparent, but the current version is not yet at the standard of a journal paper: the mathematical correction to the projection step is straightforward, but the embedding-distance issue and the confounded single-image evaluation require substantial revision and new experiments. I would not reject the work outright, but the authors need to address these points before the central claims can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper does a reasonably clean job of deriving an LLE variant that swaps Euclidean distance for an SSIM-based distance, and the supplementary material gives the gradients in detail. The kernel and out-of-sample extensions are natural and the block-wise treatment is sensible. But the central experiment does not isolate the contribution. LLISE and kernel LLISE are evaluated block-wise with majority voting, while LLE and kernel LLE appear to be run on whole 512x512 images. That means the reported advantage could come from block-level patch statistics and voting, not from SSIM distance. That is load-bearing because the paper's main selling point is that SSIM distance works better than l2 for distortion discrimination.\n\nThe derivation itself is mostly sound. The reconstruction step correctly centers the blocks so the zero-mean form of SSIM applies, and the ADMM treatment is reasonable. The embedding step is shakier: Eq. (2) expects zero-mean vectors, and the embedded y_j,i are not centered over their p coordinates; the constraint only centers the set of embedded points across images. So the objective there is a normalized squared distance rather than a true perceptual distance. The SVD projection in Eq. (20) sets singular values to n, but the constraint (1/n)Y^T Y = I implies singular values should be sqrt(n). Minor typo, but it is in the algorithm. The kernel version applies SSIM to feature-space vectors, which is conceptually odd because SSIM's constants are designed for image intensities and zero-mean in feature space is asserted via double-centering without much justification.\n\nEmpirically, the paper is thin: one Lena image, no error bars, no code, no statistical test. The out-of-sample table reports top-two vote percentages rather than full confusion information, and the mixed-distortion cases are hard to interpret. The citation pattern is fine; it cites the relevant SSIM and LLE literature without self-citation inflation.\n\nWho is this for? Researchers in image quality assessment or manifold learning who want a concrete baseline that combines structural similarity with local embedding. The idea is worth engaging with, but the current evidence does not support the headline. I would send it to peer review with a clear instruction: rerun the comparison with LLE and kernel LLE applied block-wise with the same voting protocol, ideally on multiple images and with error bars. The method is plausible, the math is mostly reproducible, and the flaw is fixable. Desk rejection would be too harsh; acceptance without a fair comparison would be too lenient.","headline":"A plausible but incremental LLE+SSIM variant whose main empirical claim is not supported by the current, uncontrolled block-wise versus whole-image comparison.","tokens_in":20161,"tokens_out":2386,"would_cite":true,"duration_ms":26444,"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":"A variant of LLE that swaps the squared Euclidean distance for an SSIM-based distance learns an image structure manifold on which distortion types separate better than LLE, except for impulse noise.","keywords":["locally linear embedding","locally linear image structural embedding","structural similarity","SSIM","image structure manifold","manifold learning","image quality assessment","out-of-sample embedding"],"falsifier":"Take the trained LLISE embeddings for a fixed block, compute the true SSIM of Eq. (1) between every pair of embedded points, and compare those values with the distance $1-\\theta_j(Y_i)$ used in Eq. (14); for any pair whose embedded coordinates are not zero-mean the two orderings will differ, showing the objective being optimized is a normalized squared distance rather than SSIM.","tokens_in":19195,"feed_emoji":"🖼️","tokens_out":13756,"duration_ms":127352,"temperature":0.7,"pith_summary":"This paper is trying to establish that a manifold-learning method built on perceptual structural similarity, rather than pixel-wise squared error, can organize images by distortion type. It introduces LLISE, a version of Locally Linear Embedding (LLE) in which the distance between image blocks is the SSIM-based distance of Eq. (2), and a kernel version operating in a feature space. In experiments on distorted versions of a single test image, a 1-nearest-neighbour classifier recognizes distortion types more accurately in the LLISE embedding than in LLE or kernel LLE for every distortion tested except impulse noise. If the claim holds, image structure manifolds of this kind give a low-dimensional representation that separates what is wrong with an image from how wrong it is.","feed_headline":"SSIM-based embedding beats LLE on distortion recognition","feed_subtitle":"In a 1-nearest-neighbour test on distorted images, LLISE separates distortion types better than LLE except for impulse noise.","key_machinery":"The load-bearing object is the SSIM distance of Eq. (2), $\\|\\check{x}_1-\\check{x}_2\\|_S := 1 - SSIM(\\check{x}_1,\\check{x}_2) = \\|\\check{x}_1-\\check{x}_2\\|_2^2 / (\\|\\check{x}_1\\|_2^2 + \\|\\check{x}_2\\|_2^2 + c)$, which is exact only for zero-mean blocks. LLISE replaces the squared Euclidean distance in both stages of LLE with this expression: first, reconstruct each image block from its $k$ neighbours by minimizing the SSIM distance of the residual; second, embed the blocks so that the same reconstruction weights are respected, again using the SSIM distance, under zero-mean and unit-covariance constraints. Because the resulting problems are nonconvex, the paper solves them with ADMM, using one gradient-descent step for the inner update and projections (normalization of the weights; centering and singular-value scaling for the embedded matrix) for the constraint. Kernel LLISE runs the same routine on double-centered kernel matrices, computing distances and gradients through the kernel trick.","core_discovery":"The paper's central claim is that replacing the $\\ell_2$ norm with the SSIM distance in both stages of Locally Linear Embedding produces a low-dimensional image structure manifold whose geometry reflects the type of structural degradation rather than the raw intensity error. For each image block, LLISE finds $k$ nearest neighbours, solves for reconstruction weights using the SSIM distance as the error, and then embeds the blocks with those weights fixed, again under the SSIM distance, subject to zero-mean and unit-covariance constraints. The kernel variant carries out the same procedure in a feature space defined by a kernel and centers the kernel matrix instead of the raw blocks. The paper reports confusion-matrix experiments in which the resulting embeddings recognize distortion types better than LLE and kernel LLE for every tested distortion except impulse noise, and out-of-sample tests in which the true distortion is usually among the top two block votes.","pith_inferences":["Because Eq. (2) presupposes zero-mean vectors while the embedded points are not centered, a useful control experiment would replace the objective with plain normalized Euclidean distance and check whether the gains persist; this would show whether the mechanism is perceptual or geometric.","The blockwise construction is not tied to single-scale SSIM; substituting multi-scale SSIM or a learned perceptual distance into the same pipeline is a natural extension that the paper leaves unexplored.","The evaluation uses one source image with distortions matched in MSE, so whether the structure manifold transfers to diverse natural images remains an open empirical question."],"forward_implications":["Images at equal MSE but different distortion types land in different regions of the LLISE manifold, making distortion type a recognizable dimension.","The kernel version extends the same separation to feature-space manifolds, so the approach is not limited to linear structure in pixel space.","Out-of-sample images, including images with mixed distortions, can be embedded by nearest-neighbour reconstruction, and the true distortion is usually among the top two block votes.","In the reported comparison, LLISE and kernel LLISE outperform LLE and kernel LLE for every tested distortion except impulse noise."],"supporting_citations":[{"why":"Establishes the motivating premise that mean squared error is a poor image-fidelity measure, which is what pushes the method toward SSIM.","marker":"[1]"},{"why":"Provides the SSIM definition and the structural versus non-structural distortion distinction that the LLISE distance is designed to capture.","marker":"[2, 3]"},{"why":"Supplies the zero-mean squared SSIM distance in Eq. (2), the exact replacement for the Euclidean norm in LLISE, and the blockwise SSIM optimization precedent.","marker":"[4, 6]"},{"why":"Defines Locally Linear Embedding, the algorithm that LLISE modifies and compares against.","marker":"[5]"},{"why":"Defines kernel LLE, the feature-space baseline that kernel LLISE extends and is compared with.","marker":"[7]"},{"why":"Provides the ADMM formulation and the projection/proximal steps used to solve the nonconvex reconstruction and embedding problems.","marker":"[8, 9]"},{"why":"Supplies the out-of-sample embedding rule by linear reconstruction, which LLISE uses for new images.","marker":"[13]"}],"fun_headline_variants":["SSIM-driven LLE variant excels in distortion separation","Swap L2 for SSIM in LLE to spot distortions better","LLISE: LLE with SSIM distance improves distortion recognition","Distortion recognition boosted by SSIM-based LLE embedding","Replacing MSE with SSIM in LLE sharpens distortion detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"In Eq. (2) the SSIM distance is derived under a zero-mean assumption, but the embedding stage applies it to embedded points that are not centered over their coordinates, so the objective being optimized is a normalized squared distance rather than the claimed SSIM distance.","fun_headline_variants_meta":{"raw":{"variants":["SSIM-driven LLE variant excels in distortion separation","Swap L2 for SSIM in LLE to spot distortions better","LLISE: LLE with SSIM distance improves distortion recognition","Distortion recognition boosted by SSIM-based LLE embedding","Replacing MSE with SSIM in LLE sharpens distortion detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000223,"raw_usage":{"total_tokens":1402,"prompt_tokens":837,"completion_tokens":565,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":453,"completion_tokens_details":{"reasoning_tokens":480}},"tokens_in":453,"tokens_out":565,"duration_ms":5506,"temperature":1.0,"reasoning_tokens":480,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:16:10.289089+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the trained LLISE embeddings for a fixed block, compute the true SSIM of Eq. (1) between every pair of embedded points, and compare those values with the distance $1-\\theta_j(Y_i)$ used in Eq. (14); for any pair whose embedded coordinates are not zero-mean the two orderings will differ, showing the objective being optimized is a normalized squared distance rather than SSIM.","supporting_citations":[{"cited_title":"IEEE signal processing magazine 26(1) (2009) 98–117","cited_arxiv_id":null,"evidence_quote":"Establishes the motivating premise that mean squared error is a poor image-fidelity measure, which is what pushes the method toward SSIM."},{"cited_title":"Science 290(5500) (2000) 2323–2326","cited_arxiv_id":null,"evidence_quote":"Defines Locally Linear Embedding, the algorithm that LLISE modifies and compares against."},{"cited_title":"EURASIP journal on Advances in signal processing (2012) 1–9","cited_arxiv_id":null,"evidence_quote":"Defines kernel LLE, the feature-space baseline that kernel LLISE extends and is compared with."},{"cited_title":"Journal of machine learning research 4 (2003) 119–155","cited_arxiv_id":null,"evidence_quote":"Supplies the out-of-sample embedding rule by linear reconstruction, which LLISE uses for new images."}],"review_version":1}