{"id":"473eff40-ea9e-476c-aef9-f3960bf4e9e9","arxiv_id":"2509.06868","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"An ALPR pipeline using selective GAN deblurring plus two YOLOv5 detectors reports 95-97% per-stage precision and a 41-point end-to-end accuracy gain on a blur-heavy test set.","lead":"This paper pairs a GAN-based blur remover with two YOLOv5 detectors to read Iranian license plates, reporting 95-97% per-stage accuracy and a roughly 40% accuracy gain when blurry plates are sharpened first. It is a practical engineering contribution for camera-based plate reading, but its headline gains are measured on a small in-house test set dominated by synthetic blur.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 40% deblur gain is measured on synthetic uniform-kernel blur; the paper itself concedes such blur is unrepresentative, so external validity on real captures is unsubstantiated.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern: the synthetic blur used for training and testing may not represent real-world blur. This is the most critical issue because the paper's headline improvement is quantified only on that synthetic distribution. The manuscript itself supplies the admission that synthetic blur datasets are not sufficiently representative (Section III.A.1), which strengthens the concern. No other issue (e.g., runtime claims, accuracy/precision labeling, or lack of comparison to prior Iranian ALPR systems) is as directly tied to the central claim's validity. The reader's CONDITIONAL verdict is appropriate: the work is internally consistent and re-implementable, but the external-validity claim needs an explicit test on real blurred data. Therefore I recommend keeping the verdict UNCHANGED, with the condition that the authors perform the real-blur evaluation or temper their claim.","tokens_in":14117,"tokens_out":4143,"duration_ms":43294,"concrete_test":"Construct a test set of genuinely blurred license-plate images, e.g., from dashcam video recorded during vehicle motion (or from the authors' own 78 videos if they contain actual motion blur). Run the full ALPR pipeline with and without the Deblur-GAN on this set and compare end-to-end accuracy. If the improvement is materially below the 41.5-percentage-point gain reported in Table 7, the central claim should be restricted to synthetic uniform-kernel blur, not real blurred captures.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim—that the Deblur-GAN preprocessor improves end-to-end ALPR accuracy by nearly 40% (from 45.80% to 87.26%, Table 7)—rests entirely on a test set of 300 images, 75% of which are blurred. The manuscript does not state the provenance of these 300 images, but the only blur-generation procedure described is in Section IV: sharp images from the ALPR dataset are convolved in OpenCV with uniform kernels of sizes 7×7 up to 19×19. The Deblur-GAN is trained on the same kind of synthetically blurred images. Thus the measured gain is in-distribution for the blur model. Critically, the paper itself acknowledges in Section III.A.1 that 'the synthetic blur datasets so generated are not sufficiently representative, specifically in a multi-source blur or blur kernel with complex patterns scenarios [22].' Real camera shake and motion blur are typically non-uniform, spatially varying, and multi-source, so the gap between training and deployment blur distributions is exactly the kind that can make a deblurring gain shrink or vanish. Without evaluation on real blurred captures—or at least on non-uniform/spatially-varying blur—the headline 'nearly 40%' improvement cannot be taken as evidence for the claimed benefit on 'blurred scenes captured by the camera as input.'","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an ALPR pipeline consisting of a selective Deblur-GAN preprocessing module, a YOLOv5 detector for license-plate localization (LPD), and a second YOLOv5 detector for character segmentation/recognition (CR). The selection mechanism uses a Laplacian-variance blur check to decide whether deblurring is applied. The authors introduce new Iranian ALPR and deblur datasets, train the components on augmented data, and report YOLOv5 LPD precision of 0.95 and CR precision of 0.976 with 0.026 s per detection stage on a Tesla T4. The central end-to-end claim is that adding the Deblur-GAN raises accuracy from 45.80% to 87.26% on a 300-image test set described as 75% blurred. A Raspberry Pi implementation is also reported.","tokens_in":14305,"tokens_out":3584,"duration_ms":36976,"significance":"The selective-deblur design is sensible and the public release of Iranian ALPR data is a useful community contribution. If the 40% improvement were validated on realistic blurred captures, the paper would demonstrate a practically valuable preprocessing strategy for ALPR. The YOLOv5 speed numbers are attractive for embedded use. However, the headline gain is measured only on synthetic uniform-kernel blur of the same type used for training, on a small test set without uncertainty quantification, and the real-time claim is based on the non-deblurred path. These gaps currently prevent the central claims from being accepted as stated.","major_comments":[{"comment":"The 40% improvement from 45.80% to 87.26% is measured on 300 images described only as '75% blurred'. The provenance of these images is not stated, but the only blur-generation procedure in the paper is Section IV: convolving sharp images with uniform OpenCV kernels of size 7x7 to 19x19. The Deblur-GAN was trained on exactly this synthetic degradation. Section III.A.1 itself concedes, citing [22], that such synthetic blur is 'not sufficiently representative, specifically in a multi-source blur or blur kernel with complex patterns'. Therefore the reported gain is an in-distribution result for the trained blur model and does not establish performance on real camera-shake or motion blur. Please evaluate on real blurred captures or at least on non-uniform/spatially-varying blur, and report per-blur-kernel and per-blur-level results.","section":"Table 7 / Section V.A"},{"comment":"The central quantitative claim rests on 300 end-to-end images (225 blurred), with no confidence intervals, no repeated runs, and no per-condition breakdown. The difference 45.80% vs 87.26% is an absolute improvement of 41.46 percentage points, which is not 'nearly 40%' in relative terms (relative improvement is ~90%). Please report uncertainties (e.g., Wilson intervals or bootstrap), a significance test, and a breakdown by blur level and by sharp vs blurred subsets. Without this, the headline accuracy gain is not statistically grounded.","section":"Table 7 / Abstract"},{"comment":"Table 5 states 80/10/10 train/validation/test splits for the YOLO models, but Table 6 reports 3,752 test images for LPD and 1,525 for CR. These do not match 10% of the stated dataset sizes (39,966 and 28,854). The source and composition of these test sets should be described precisely. In addition, the Abstract's '95% and 97% accuracy' figures correspond to the Precision column of Table 6 (0.95 and 0.976), not to an accuracy metric. Please clarify which metric is being reported and, if plate-level accuracy is intended, compute it explicitly.","section":"Table 5 vs Table 6"},{"comment":"The real-time claim is based on the non-deblurred path: Table 7 gives 0.052 s end-to-end without deblurring but 2.052 s with Deblur-GAN. Table 8 compares only the 0.052 s figure against prior work, omitting the 2.052 s selective-deblur pipeline that would be invoked for blurred inputs. The paper should report the end-to-end latency of the full selective pipeline, including the blur check and deblurring when triggered, and discuss whether the system remains real-time under the stated 75%-blur test condition.","section":"Table 7 / Table 8 / Abstract"}],"minor_comments":[{"comment":"The blur-check threshold is described as a 'carefully selected user-set threshold', but no value, selection procedure, or sensitivity analysis is given. Since the selective behavior depends on it, at least a brief description of its calibration should be added.","section":"Section III.B.1"},{"comment":"Typographical errors and formatting issues: 'Sumsung' should be 'Samsung'; several cells contain 'NA' repeated as 'NANANP'; superscript footnote markers are unexplained. Also, the table reports 'Blur filter size 7×7,9×9,...,19×9' but the text says '7×7 up to 19×19'.","section":"Table 3"},{"comment":"The dataset link is given only as a GitHub URL [36] with no version or DOI; if the data is publicly released, please provide a stable identifier and clarify the licensing.","section":"References"},{"comment":"The figure caption says 'without/with Deblur-GAN (left/right)', but the ordering is ambiguous from the text. Please make the caption explicit and ensure all subfigures are labeled consistently.","section":"Figure 10"}],"recommendation":"major_revision","confidential_remarks":"The core engineering contribution—selective GAN deblurring ahead of YOLOv5 ALPR—is plausible and the dataset release is valuable. However, the headline 40% gain is based on an in-distribution synthetic evaluation with a small test set and no statistical grounding. The authors should be asked to add real-blur or non-uniform-blur validation, confidence intervals, and a corrected reporting of speed and accuracy metrics before the paper can be accepted. The issues are substantial but within the scope of additional experiments and reporting."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a legit engineering paper, not a research breakthrough, but it ships two new public datasets and a clean gate-then-deblur preprocessor for ALPR. The headline 40% end-to-end gain under blur is measured on synthetic uniform-kernel blur that the authors themselves concede is not representative of real blur, so don't quote that number for real-world systems.\n\nWhat's new: the Iranian ALPR and Deblur datasets (released on GitHub), the selective blur-gating design, and a measured end-to-end improvement from 45.80% to 87.26% on a 300-image test set. The per-stage numbers (0.95 LPD precision, 0.976 CR precision) are plausible for off-the-shelf YOLOv5 on a moderate dataset. The paper is internally consistent and the method is re-implementable.\n\nThe main soft spot is external validity. The blur used for training and testing is generated by convolving sharp images with uniform kernels (7x7 to 19x19) in OpenCV. The test set is 75% blurred and comes from the same synthetic pipeline. The paper itself cites Schuler et al. [22] for the point that synthetic blur is not representative of multi-source or complex-pattern blur, yet that is exactly what real camera shake and motion blur often look like. No evaluation on real blurred captures or on non-uniform blur is reported, so the 40% gain is in-distribution and may shrink or vanish in deployment.\n\nSmaller issues: no error bars or repeated-run variability; the abstract calls precision 'accuracy' (Table 6 reports Precision and Recall); the 0.026s real-time claim applies only to the YOLO stages, not the 2.052s end-to-end path when deblurring is triggered; and the comparison to prior Iranian ALPR systems is only on speed, not accuracy.\n\nThe reader's conditional verdict matches my own. This is a useful engineering contribution that should go to peer review, but the authors need to test on real blurred images or at least on a non-uniform blur model, report the blur-detection threshold and test-set provenance, add error bars, and fix the abstract. With those changes it would be a solid contribution to the ALPR literature.\n\nWho should read it: people working on Iranian plates, and anyone designing selective preprocessors for detection pipelines. It deserves a serious referee.","headline":"Useful engineering report with two released datasets and a gated deblurrer, but the headline 40% gain is measured on in-distribution synthetic blur and should not be trusted for real captures.","tokens_in":14971,"tokens_out":3541,"would_cite":false,"duration_ms":31999,"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 selective GAN deblurring step before YOLOv5 detection raises end-to-end license-plate recognition accuracy from 45.80% to 87.26% on blurred images.","keywords":["automatic license plate recognition","GAN deblurring","YOLOv5","object detection","character recognition","selective preprocessing","Iranian license plates","real-time detection"],"falsifier":"Take real blurred license-plate images from dashcam or handheld video, run the pipeline with and without the Deblur-GAN, and compare end-to-end accuracy; if the gap is much below the reported 41.46 percentage points, the claim is specific to synthetic blur.","tokens_in":13901,"feed_emoji":"🚗","tokens_out":4181,"duration_ms":37786,"temperature":0.7,"pith_summary":"This paper aims to show that a complete automatic license-plate recognition (ALPR) pipeline can get much better accuracy on blurred images by adding a selective GAN-based deblurring step before a YOLOv5 detector, and that the YOLOv5 stages are fast enough for near-real-time use. The authors report that the deblur step raises end-to-end accuracy from 45.80% to 87.26% on a test set where three-quarters of the images are blurred. They also report 95% license-plate detection accuracy and 97% character-recognition accuracy, at 0.026 seconds per detection stage on a Tesla T4 GPU, and a Raspberry Pi implementation at 0.24 seconds total. If these numbers hold, the main value is a practical recipe: blur only when needed, and let a fast one-stage detector handle plate and character localization.","feed_headline":"Selective GAN deblurring lifts ALPR accuracy from 46% to 87%","feed_subtitle":"YOLOv5 stages run at 0.026 s; selective GAN deblurring nearly doubles end-to-end accuracy on blurred plates.","key_machinery":"The load-bearing mechanism is the selective-deblur module: a Laplacian kernel computes edge intensity, a variance threshold decides whether the image is blurred, and only blurred images pass through the Deblur-GAN, a multi-scale Deep Multi-scale CNN generator (three resolution scales, residual blocks, up-convolution concatenation) with a discriminator. The two YOLOv5 heads then carry localization (one class: license plate) and character recognition (44 classes for Iranian plates), with anchor boxes and output head sizes 18 and 147 respectively.","core_discovery":"The central claim is that a selective deblurring pre-processor based on a multi-scale GAN, gated by a Laplacian-variance blur check, is the difference between a failing and a working ALPR system on blurred captures. Without the deblurrer, the YOLOv5 pipeline recognizes 45.80% of plates in a heavily blurred test set; with it, accuracy rises to 87.26%. The paper further claims that YOLOv5 is a good backbone choice for both license-plate detection and character segmentation+recognition because it detects small objects well and runs in 0.026 s per stage, and that the generated public datasets, including augmented weather conditions and synthetic uniform blur kernels, are representative enough to","pith_inferences":["Because the deblurring benefit is measured entirely on images blurred with synthetic uniform kernels, deploying on real camera-shake blur—often non-uniform and multi-source—could yield a smaller gain; a real-blur benchmark would settle this.","The 2.052 s deblurring cost suggests the selective gate matters: on mostly-sharp video streams, skipping the GAN keeps the pipeline at 0.052 s, so the threshold could be tuned per camera.","The same selective pre-processing pattern—a cheap blur detector followed by an expensive restoration network—could generalize to other small-object recognition tasks such as traffic-sign or pedestrian detection.","A cheaper or distilled deblur network might retain most of the accuracy gain while reducing the 2-second penalty on blurred frames, making the selective module usable in closer to real-time."],"forward_implications":["End-to-end recognition on heavily blurred plates jumps from 45.80% to 87.26% when the selective GAN deblurrer is active.","YOLOv5 LPD and CR each detect in 0.026 seconds on a Tesla T4, making the non-blurred path near real-time.","The same pipeline reaches 0.24 seconds total on a Raspberry Pi-4B, suggesting portability to dashcam-like devices.","Publicly released ALPR and deblur datasets enable others to retrain and benchmark Iranian-plate ALPR systems under varied weather and blur.","Against prior Iranian ALPR systems, the reported per-image runtime of 0.052 s is lower than the compared baselines (0.120-0.356 s)."],"supporting_citations":[{"why":"Supplies the Deep Multi-scale CNN architecture the Deblur-GAN generator is built on.","marker":"[35]"},{"why":"The YOLOv5 detector implementation used for both license-plate detection and character recognition.","marker":"[31]"},{"why":"Baseline end-to-end ALPR on the CCPD dataset that motivates the comparison and the discussion of dataset comprehensiveness.","marker":"[11]"},{"why":"Raises the limitation that synthetic blur datasets are not representative, which the paper's deblur dataset attempts to address.","marker":"[22]"},{"why":"The publicly released ALPR and Deblur datasets used to train the models.","marker":"[36]"},{"why":"imgaug tool used to generate the augmented weather-condition training images.","marker":"[37]"},{"why":"Previous Iranian ALPR system used as a runtime comparison baseline.","marker":"[14]"},{"why":"YOLOv3 baseline against which YOLOv5 LPD/CR performance is compared.","marker":"[29]"}],"fun_headline_variants":["Blur-checked GAN preprocessing lifts plate recognition to 87%","YOLOv5 + selective GAN: 87% accuracy on blurred plates","Deblur only when needed: ALPR jumps 46% to 87% on blur","Selective deblurring boosts ALPR from 46% to 87% on blur","GAN deblurrer boosts blurred plate recognition to 87%"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The deblurring gain is measured on test images from the same synthetic uniform-blur pipeline used for training, so the assumption that this blur represents real camera motion is what carries the 40% improvement claim.","fun_headline_variants_meta":{"raw":{"variants":["Blur-checked GAN preprocessing lifts plate recognition to 87%","YOLOv5 + selective GAN: 87% accuracy on blurred plates","Deblur only when needed: ALPR jumps 46% to 87% on blur","Selective deblurring boosts ALPR from 46% to 87% on blur","GAN deblurrer boosts blurred plate recognition to 87%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001752,"raw_usage":{"total_tokens":6833,"prompt_tokens":900,"completion_tokens":5933,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":644,"completion_tokens_details":{"reasoning_tokens":5826}},"tokens_in":644,"tokens_out":5933,"duration_ms":46075,"temperature":1.0,"reasoning_tokens":5826,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T22:58:19.819072+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take real blurred license-plate images from dashcam or handheld video, run the pipeline with and without the Deblur-GAN, and compare end-to-end accuracy; if the gap is much below the reported 41.46 percentage points, the claim is specific to synthetic blur.","supporting_citations":[{"cited_title":"Deep multi-scale convolutional neural network for dynamic scene deblurring,","cited_arxiv_id":null,"evidence_quote":"Supplies the Deep Multi-scale CNN architecture the Deblur-GAN generator is built on."},{"cited_title":"Jocheret al., “Yolov5.” [Online]","cited_arxiv_id":null,"evidence_quote":"The YOLOv5 detector implementation used for both license-plate detection and character recognition."},{"cited_title":"Towards end-to-end license plate detection and recognition: A large dataset and baseline,","cited_arxiv_id":null,"evidence_quote":"Baseline end-to-end ALPR on the CCPD dataset that motivates the comparison and the discussion of dataset comprehensiveness."},{"cited_title":"Learning to deblur,","cited_arxiv_id":null,"evidence_quote":"Raises the limitation that synthetic blur datasets are not representative, which the paper's deblur dataset attempts to address."},{"cited_title":"Dataset,","cited_arxiv_id":null,"evidence_quote":"The publicly released ALPR and Deblur datasets used to train the models."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"imgaug tool used to generate the augmented weather-condition training images."},{"cited_title":"A Robust Deep Learning Approach for Automatic Iranian Vehicle License Plate Detection and Recognition for Surveillance Systems,","cited_arxiv_id":null,"evidence_quote":"Previous Iranian ALPR system used as a runtime comparison baseline."}],"review_version":1}