{"id":"7fa184ef-002a-4334-9c70-2b9d798788b5","arxiv_id":"2507.00527","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":2.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A course project applies standard thresholding, median filtering, frequency masking, and Wiener filtering to suppress aliasing in naked-eye 3D display images, with only qualitative before-and-after photos as evidence.","lead":"The authors tested frequency-domain masking and Wiener restoration to reduce jagged, lens-induced distortions in naked-eye 3D display images. The results look smoother in sample photos, but the paper provides no measurements, no comparison to existing filters, and no code, so its practical value is unproven.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim rests on an unverified single degradation function: no H(u,v) is ever shown or validated, and the implemented Wiener filter uses hand-tuned phase combination and radial masking, so the reported success may reflect fitting to one test pattern.","rationale":"For the central claim to hold, there must be an actual degradation function H that can be identified from a clean/degraded pair and that transfers to other images. Neither is demonstrated. The closest implementation detail, the Wiener pseudocode, uses `spectrum_magnitude * K` and phase `2*phase1 - phase2`; these are not the inverse of any measured filter response. Consequently, the before/after figures cannot distinguish successful restoration from an aggressive frequency mask tuned to one 80-degree line pattern. The paper even acknowledges (Section V) that text images lose more information in preprocessing than lens distortion causes, which bounds the generality of the method. The reader's weakest_assumption is correct as far as it goes: the single-H LSI model is asserted, not validated. I would sharpen it: the authors do not even exhibit a concrete H, so the model cannot be checked. A targeted experiment estimating H from multiple pairs and testing fixed-H restoration on held-out images would settle whether the claim has any real basis. Because no such evidence is present, the REJECT verdict is unchanged; this is not an internal inconsistency but a missing derivation and missing validation.","tokens_in":6014,"tokens_out":3833,"duration_ms":47475,"concrete_test":"Re-derive the transfer function from the Section IV original/degraded pair as H(u,v) = FFT(degraded)/FFT(original) with a small regularization floor, and display it. Then (i) compare H estimated from at least two different pairs with different line widths or positions; (ii) apply the inverse or radial-Wiener filter with that fixed H to an unseen test image captured through the same lens array; (iii) score restoration with a quantitative metric such as PSNR/SSIM or line-edge distance against the known ideal. If H differs substantially between pairs, or if the fixed-H restoration fails on unseen images in a way that per-image tuned cutoffs do not, then the single-LSI assumption underlying the claim is falsified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV promises a degradation function estimated from original/degraded pairs, and the conclusion asserts that 'the degradation function and radial Wiener filtering are very effective in dealing with aliasing.' Yet no H(u,v) is ever written down, plotted, or numerically specified. The Wiener-filter pseudocode combines phase terms as `2*phase1 - phase2` with a scaled magnitude and a radial low-pass (cutoff 17), followed by custom binarization at threshold 140; this is a hand-tuned phase/magnitude heuristic, not an inverse or Wiener filter constructed from a measured transfer function. The governing assumption of the whole restoration branch is that the lens array's effect is one linear shift-invariant convolution plus noise, so a single H applies to all images and regions. The paper supplies no validation of this model; Section III itself says that around nodes 'pixel values ... get close to the grey values of noise' due to photographing limitations, which is inconsistent with a stationary LSI blur. With neither an extracted H nor any quantitative evaluation (PSNR, SSIM, edge error, or comparison against baseline anti-aliasing filters), the central claim is not supported beyond a few subjective before/after line-pattern images.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a post-processing pipeline to reduce aliasing and jaggedness that appear when line patterns are viewed through a naked-eye 3D display with a columnar lens array. The authors identify an approximately 80-degree tilt direction as the worst case, preprocess images with global thresholding (T=60) and median filters (7x7), and then test two frequency-domain strategies: masking a parallelogram region of the magnitude spectrum, and 'degradation function filtering' using inverse, radial, and Wiener-type filtering. The paper claims to locate the root cause of node blurring and to have extracted the degradation function of the lens array, but it does not report a quantitative evaluation or an actual specification of H(u,v), and it relies on a handful of before/after photographs of one test pattern.","tokens_in":6219,"tokens_out":5613,"duration_ms":67726,"significance":"If the central claims were established, the work would offer a simple software-only method for mitigating lens-induced artifacts in autostereoscopic displays and a step toward hardware-level insight via an estimated degradation function. The paper has some genuine ingredients: the observation of the 80-degree/170-degree spatial-frequency relation is concrete, the pipeline is described well enough to be re-implemented from the pseudocode, and the authors are explicit about several limitations (e.g., small text images are not handled well, Section V). However, the manuscript's significance is currently not supported: there are no quantitative metrics, no comparison with the cited anti-aliasing filters or any baseline, and the central degradation function is never shown or validated.","major_comments":[{"comment":"The central claim that 'the degradation function and radial Wiener filtering are very effective' rests on an estimated H(u,v), but no H(u,v) is ever written down, plotted, tabulated, or numerically specified in the manuscript. Section IV says the function is derived by comparing original and degraded images, yet the described Wiener implementation combines two phase spectra as 2*phase1 - phase2, multiplies the magnitude by a hand-tuned scaling factor K, and applies a radial low-pass mask with cutoff 17. This is a phase/magnitude heuristic, not an inverse filter or Wiener filter constructed from a measured transfer function. Without the extracted H, the paper cannot support its central claim that the root cause of node blurring has been located.","section":"Section IV and pseudocode (Wiener Filter, Steps 1-6)"},{"comment":"The restoration branch assumes the lens array's effect is a single linear, shift-invariant convolution plus additive noise, so that one degradation function H applies to all images and regions. This model is not validated anywhere. The text itself notes that around the 'nodes' the pixel values approach the grey values of noise due to photographing limitations, which indicates a non-stationary, signal-dependent component that an LSI model cannot represent. If the LSI assumption fails, inverse filtering and Wiener filtering will not reliably recover clean images, and the reported success may be an artifact of tuning to one test pattern.","section":"Section III (Image Preprocessing) and Section IV"},{"comment":"There is no quantitative evaluation of any kind: no PSNR, SSIM, edge error, contrast measure, or statistical comparison, and no comparison with standard anti-aliasing filters such as the references cited in the paper. The conclusion that the methods 'are both capable of eliminating degradation' is supported only by visual inspection of a small number of before/after line-pattern photographs. As the parameters (T=60, 7x7 median, cutoff=17, custom binarization threshold=140, K, and parallelogram vertices) are all hand-tuned on this test image, the reported effectiveness cannot be separated from overfitting to the single example.","section":"Figures 11-21 and Section VI"},{"comment":"The manuscript contains an internal inconsistency between its limitation statement and its conclusion. Section V concedes that for small text images, 'the image loss caused by binarization and median filtering far exceeds the degradation caused by the cylindrical lens distortion,' yet Section VI concludes that the proposed methods can 'eliminate degradation and improve displaying performance' without stating this restriction. Either the conclusion must be explicitly scoped to line-pattern images, or a text-specific procedure must be demonstrated; as written, the general claim is not supported by the paper's own evidence.","section":"Section V (Discussion)"}],"minor_comments":[{"comment":"Two sections are labeled 'III' (Image Preprocessing and Frequency-Domain Filtering), and Section IV appears after the second 'III'; the figure numbering is also inconsistent, with Figures 8, 19, and 20 used twice.","section":"Section numbering and figures"},{"comment":"There are repeated typos: 'Winer' for 'Wiener' in Figures 20-21 and Section VI; 'naked eyes' for 'naked-eye'; 'degenerate function' where 'degradation function' is meant in the Abstract and Introduction.","section":"Throughout"},{"comment":"The text contains a broken sentence: 'Wiener filtering is especially suitable for restoring degraded images in high-' followed by a line break and 'noise scenarios'; this should be corrected.","section":"Section IV"},{"comment":"The pseudocode functions `calculate_scaling_factor`, `extract_phases`, and `generate_radial_lowpass_filter` are never defined; their behavior is essential for reproducing the method.","section":"Pseudocode (Wiener Filter)"},{"comment":"The bibliography contains only two entries and no references for standard inverse filtering, Wiener filtering, or Fourier-based deconvolution; the related-work discussion is therefore minimal.","section":"References"}],"recommendation":"reject","confidential_remarks":"This is a course-project-style manuscript whose claims outrun its evidence. Even a major revision would require adding a complete evaluation methodology, extracting and validating H, testing on multiple images and patterns with quantitative metrics and baseline comparisons, and clearly scoping the conclusion. In my view that is beyond a normal revision cycle for this venue, though the underlying problem is legitimate. The paper's self-report that it was developed without online resources is not relevant to scientific assessment."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper is a course project report, and it reads like one. The genuinely useful observation is that lens-induced aliasing in this naked-eye 3D display is worst around an 80-degree tilt, and the paper documents a plausible two-branch processing pipeline (direct frequency masking and restoration-style filtering) with pseudocode and honest acknowledgment of limitations, especially for text images. That is real work, and the authors do not hide that the parameters are hand-tuned after trials. If this were submitted as a lab report, it would probably pass.\n\nBut as a scientific preprint, the central claim is not supported. The stress-test concern holds up: Section IV promises a degradation function estimated from original/degraded pairs, and the conclusion says that function plus radial Wiener filtering is \"very effective.\" No H(u,v) is ever written down, plotted, or numerically specified. The pseudocode for the \"Wiener filter\" is not a Wiener filter in any standard sense—it combines phase terms from two images with a scaling factor and then applies a radial low-pass cutoff at 17. That is a hand-built frequency-domain heuristic, not a filter derived from a measured or estimated transfer function. The paper also never validates the linear shift-invariant assumption; Section III itself notes that pixel values around nodes are corrupted by photographing limitations, which is inconsistent with a stationary LSI blur. There are no quantitative metrics, no comparison to the cited anti-aliasing filters (refs [1,2]), and no actual code or data shipped—only pseudocode. The before/after photos are subjective, and the parameters appear to be selected to make one test pattern look clean.\n\nWhat is not in the paper is more damning than what is: no PSNR, SSIM, edge error, or even a second test pattern to show generalization. The \"root cause\" language in the conclusion is inflated. The paper demonstrates familiarity with basic Fourier-domain reasoning but does not establish the engineering claim it asserts.\n\nWho is this for? Only someone interested in what a competent student team can produce in a display-image processing course. It does not deserve a serious referee: the evidence is too thin and the methods too ad hoc. I would not cite it, and I would not bring it to reading group. If the authors continue, they should add quantitative evaluation, release actual code and data, and show the estimated H—or stop calling it a degradation function.\n\nRecommendation: desk reject. The work is earnest but not yet a scientific contribution.","headline":"Student course report that hand-tunes standard filters on one test pattern and never actually shows the degradation function it claims to extract.","tokens_in":6790,"tokens_out":1547,"would_cite":false,"duration_ms":20507,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Lens-induced jaggies in naked-eye 3D displays can be removed by estimating a single degradation function and applying radial Wiener filtering in the frequency domain.","keywords":["autostereoscopic display","naked-eye 3D display","columnar lens array","anti-aliasing","frequency-domain filtering","Wiener filtering","degradation function estimation","image restoration"],"falsifier":"Photograph a fresh 80-degree line pattern through the display, estimate the degradation function from a different region or image, apply the inverse or radial Wiener filter, and compare the restored image with the ideal: if node-like artifacts remain in some regions or ringing appears where the original was clean, the single-function assumption fails.","tokens_in":5757,"feed_emoji":"🖥️","tokens_out":6340,"duration_ms":67015,"temperature":0.7,"pith_summary":"This paper tries to establish that the jagged, node-like distortion seen when viewing images through a cylindrical-lens naked-eye 3D display has a single identifiable cause: the lens array's tilt of roughly 80 degrees. It argues that this distortion can be removed in software by first masking the aliasing frequencies in the Fourier domain and then applying radial Wiener filtering with a degradation function estimated from an ideal and a photographed image. If the claim holds, a post-processing pipeline rather than a hardware redesign can clean up images and text before they are shown on such displays. The paper reports that the degradation function and radial Wiener filtering are very effective in dealing with aliasing, while acknowledging that small text remains a difficult case.","feed_headline":"One blur estimate removes jaggies from naked-eye 3D displays","feed_subtitle":"Frequency masking plus radial Wiener filtering restore distorted 80-degree line images in software.","key_machinery":"The central object is the degradation function $H(u,v)$ of the columnar lens array, estimated in the frequency domain from the ratio of a photographed degraded image to the ideal pattern. The argument's second mechanism is the geometric fact that a spatial distortion at 80 degrees appears as a perpendicular line at roughly 170 degrees in the Fourier magnitude spectrum, which lets the algorithm mask a parallelogram-shaped frequency region to suppress aliasing. The third mechanism is radial Wiener filtering, which combines the estimated degradation function with noise statistics and a low-pass cutoff of 17 to restore the image while limiting high-frequency noise.","core_discovery":"The central claim is that the blurring and 'nodes' that appear along straight lines on the display are caused by the columnar lens array's 80-degree tilt, and that the full degradation can be captured by a single degradation function in the frequency domain. Using a photographed binary pattern, the authors estimate this function and restore degraded images by inverse filtering and by radial Wiener filtering with a cutoff frequency of 17, followed by custom binarization at threshold 140. They conclude that frequency-domain masking removes the disturbing nodes, and that radial Wiener filtering is 'very effective in dealing with aliasing'.","pith_inferences":["Assuming the single-function model holds, the same radial Wiener filter could be implemented as a real-time shader in the display driver, making the correction invisible to content creators.","The paper's own observation that pixel values around nodes are corrupted by photographing limitations suggests the degradation may be spatially varying; a residual analysis across different regions would test whether a single $H(u,v)$ is sufficient.","The suggested spatial-domain slicing-and-shifting correction could be combined with the frequency-domain estimate to handle text, which the paper identifies as a failure case.","A natural next experiment is to rotate arbitrary line angles into the 80-degree orientation, restore them, and rotate back, which would generalize the method beyond a single tilt angle."],"forward_implications":["If the single degradation function is valid, the same estimated filter can be reused on other images without re-photographing the display.","Frequency-domain masking alone removes the node artifacts from straight-line patterns, making it a cheap first-pass anti-aliasing step.","Radial Wiener filtering can be applied before an image is sent to the 3D display, improving what the viewer sees rather than only post-processing a photo.","The method works for line patterns of different thicknesses, indicating it is not tuned to one specific test image.","Small text remains outside the current method's reach because binarization and median filtering destroy more detail than the lens distortion."],"supporting_citations":[],"fun_headline_variants":["Frequency masking and Wiener filter fix naked-eye 3D jaggies","Single blur model cleans up naked-eye 3D display distortion","Blur estimate reverses jaggies on tilted-lens 3D screens","Radial Wiener filter strips jaggies from 3D display lines","Naked-eye 3D jaggies fixed via frequency-domain masking"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole method rests on the assumption that the lens array blurs every image in the same fixed way plus random noise, so a single degradation function estimated from one pair of pictures works everywhere.","fun_headline_variants_meta":{"raw":{"variants":["Frequency masking and Wiener filter fix naked-eye 3D jaggies","Single blur model cleans up naked-eye 3D display distortion","Blur estimate reverses jaggies on tilted-lens 3D screens","Radial Wiener filter strips jaggies from 3D display lines","Naked-eye 3D jaggies fixed via frequency-domain masking"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000189,"raw_usage":{"total_tokens":1225,"prompt_tokens":721,"completion_tokens":504,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":337,"completion_tokens_details":{"reasoning_tokens":409}},"tokens_in":337,"tokens_out":504,"duration_ms":5296,"temperature":1.0,"reasoning_tokens":409,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:12:48.388220+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Photograph a fresh 80-degree line pattern through the display, estimate the degradation function from a different region or image, apply the inverse or radial Wiener filter, and compare the restored image with the ideal: if node-like artifacts remain in some regions or ringing appears where the original was clean, the single-function assumption fails.","supporting_citations":[],"review_version":1}