{"id":"4fb103ad-9a05-4deb-8b4a-a01a947a32d3","arxiv_id":"2507.02354","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A YOLOv8n variant reports 92.7% mAP on shrimp disease detection with 2.1M parameters, but the small private dataset and single-run results make the gains unverified.","lead":"This paper modifies YOLOv8n with a reparameterized detection head, a multi-scale convolution module, and an attention mechanism to build a lightweight shrimp disease detector. The reported 92.7% mAP on a private 1047-image dataset with 2.1M parameters sounds useful, but the evidence is thin without released code, data, or error bars.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported 3-point mAP gain rests on a single-run 9:1 split of 1047 images, and Table 3 exp7 exactly duplicates the baseline, so the ablation does not reliably establish that the modules cause the improvement.","rationale":"The reader's CONDITIONAL verdict is appropriate. The paper's architectural changes are plausible and the parameter counts are internally consistent, but the experimental evidence for the central claim is not yet sufficient. The exact duplicate of baseline values in exp7 is the sharpest red flag: either the table contains an error, in which case all numbers are in doubt, or the modules genuinely cancel in a way that contradicts the additive interpretation. The small private dataset and single-run protocol compound the problem, since a 3-point mAP difference may be within noise. The URPC2020 experiment is also under-specified; without stating the fine-tuning protocol and class-head adaptation, the generalization claim cannot be assessed. These issues are addressable with multi-seed reruns, code/data release, and a precise experimental protocol, so a conditional acceptance requiring those fixes is the right recommendation; I do not see a reason to move the verdict.","tokens_in":8913,"tokens_out":5323,"duration_ms":59752,"concrete_test":"Re-run experiments 1, 7, and 8 from Table 3 with five random seeds each on the same 9:1 split and report mean±std of mAP@0.5; then check whether the +3.0-point gap between experiment 8 and the baseline exceeds one standard deviation. Also recompute experiment 7 from the training logs to verify whether it truly reproduces the baseline or the row was copied. If the gap is within noise or exp7 is a copy error, the headline improvement is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that RLDD, C2f-EMCM, and SegNext_Attention improve mAP@0.5 from 89.7% to 92.7% while cutting parameters by 32.3%. For this claim to hold, the evaluations must be fair and the ablation must show each module's contribution. Both are insecure. First, the dataset is 1047 images split 9:1 into train/val with no separate test set, and all results are from a single run. With roughly 105 validation images, a 3-point mAP difference can be within sampling noise, and no variance or multi-seed statistics are reported. Second, Table 3, experiment 7 (EMCM + SegNext_Attention) reports Precision 78.4% and mAP@0.5 89.7%, exactly equal to experiment 1 baseline. Since EMCM alone adds 2.4 points and SegNext_Attention alone adds 0.8 points, an exact cancellation to the same value is implausible and unexplained; it suggests a copy-paste error or a pathological interaction that undermines the compositional interpretation of the final model. Third, the URPC2020 generalization experiment does not state whether the model was fine-tuned on URPC2020 or evaluated zero-shot, nor how the class-head mismatch was handled, so the 87.2% mAP generalization claim is not reproducible from the text.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes three modifications to YOLOv8n for shrimp disease detection: a reparameterized lightweight detection head (RLDD), an efficient multi-scale convolution module (C2f-EMCM), and an improved SegNext_Attention self-attention module. On a self-constructed dataset of 1047 images containing three shrimp disease classes, the authors report mAP@0.5 of 92.7% with 2.1M parameters, compared with 89.7% and 3.1M for the YOLOv8n baseline, and 87.2% mAP on the URPC2020 dataset. The paper's evidence is entirely experimental: an eight-row ablation study, comparisons with several YOLO variants and lightweight backbones, qualitative detection visualizations, and a two-model generalization experiment on URPC2020.","tokens_in":9219,"tokens_out":3921,"duration_ms":44399,"significance":"If the reported numbers are reproducible, the combination of a 3.0-point mAP gain with a 32.3% parameter reduction is practically useful for edge deployment in aquaculture monitoring, and the improved URPC2020 result would be an encouraging sign of cross-domain robustness. The paper also has a reasonable structure: each proposed component is described, an ablation is attempted, and the final model is compared with six alternatives. However, the empirical support is currently too thin to certify the central claim. The results rest on a single run over one small private train/validation split, the ablation contains an unexplained exact duplication of the baseline, and the URPC2020 protocol is not specified. These gaps make the headline accuracy and efficiency claims unverifiable from the manuscript as written.","major_comments":[{"comment":"The entire accuracy claim rests on a single training run with a 9:1 random split of 1047 images, and the manuscript reports no separate test set, no variance across random seeds, and no confidence intervals. With roughly 105 validation images, the reported 3.0-point mAP gain and 9.1-point precision gain over YOLOv8n could be within sampling noise. Please report mean and standard deviation over at least three seeds, and preferably use a fixed held-out test split that is never used for model selection.","section":"§3.1.2 / §3.2.1"},{"comment":"Experiment 7 (EMCM + SegNext_Attention) reports Precision 78.4% and mAP@0.5 89.7%, exactly equal to the baseline experiment 1, even though experiments 3 and 4 individually improve both metrics. If this exact duplication is not a copying or recording error, it needs a mechanistic explanation; as presented, it undermines the compositional interpretation of the final model and casts doubt on the reliability of the ablation table.","section":"§3.2.1, Table 3"},{"comment":"The generalization experiment on URPC2020 does not state whether the model was fine-tuned on URPC2020 or evaluated zero-shot, how many training epochs and what hyperparameters were used, or how the detection head was adapted to the different class set (starfish, sea urchin, sea cucumber, etc.). Without this information, the reported 87.2% mAP and the 4.1-point improvement over YOLOv8n are not reproducible and cannot be interpreted as evidence of domain generalization.","section":"§3.2.4"},{"comment":"The paper repeatedly claims reduced 'computational complexity' and superiority in 'model size,' but Table 4 reports only parameter counts and does not include FLOPs, inference time, or model file size. Please add these quantities for all compared models; otherwise the efficiency part of the central claim is only partially supported.","section":"§3.2.2 / Abstract"}],"minor_comments":[{"comment":"Equation (6) appears to define AP per class but refers to 'n' as the number of classes; please clarify the interpolation formula and use consistent notation for the class index and the number of recall intervals.","section":"§3.1.3"},{"comment":"The experimental environment lists 'PyTorch 1.19.0,' which is not a released version of PyTorch; please correct this or specify the exact version used.","section":"§3.1.1"},{"comment":"The text jumps from Table 1 to Table 3; there is no Table 2. Please renumber the tables or add the missing table.","section":"General"},{"comment":"Equations (1) and (2) are poorly typeset: the summation argument, the convolution notation, and the element-wise multiplication symbol are unclear. Please rewrite these equations with standard mathematical notation.","section":"§2.3"},{"comment":"The visualization in Figure 7 is described as showing improvements on specific categories, but no per-class AP values are reported for URPC2020; please include per-class results to substantiate the qualitative claims.","section":"§3.2.4"}],"recommendation":"major_revision","confidential_remarks":"The core numerical claim is plausible but not yet established. I would ask the editor to require multi-seed results, a clarified URPC2020 protocol, and a check of Table 3 experiment 7 before any acceptance decision. If the exact duplicate in the ablation is a data-entry error, that should be corrected; if it is not, the authors need to explain the non-additivity of their modules."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a routine YOLOv8n modification (reparameterized head, grouped multi-scale convolution, SegNext attention) with a genuinely new dataset for shrimp disease detection. The reported 92.7% mAP on that dataset is plausible, but the experimental support is thinner than the abstract suggests.\n\nWhat is actually new is the dataset: 1047 images, three disease classes, annotated and used to compare several YOLO variants. That is a useful contribution for a niche but real application. The architecture is a sober recombination of known tricks; nothing here will change object detection, but the parameter reduction (32.3%) with maintained mAP is a reasonable engineering result. The paper is clearly written, and the URPC2020 generalization test is a good instinct.\n\nThe soft spots are all in the evaluation. One run on a 9:1 split of 1047 images, with no separate test set and no variance, means the 3-point mAP gain could be sampling noise. More worrying is Table 3: experiment 7 (EMCM + SegNext_Attention) reports exactly the baseline precision (78.4%) and mAP (89.7%). Exact cancellation to the same two decimals is implausible; this looks like a copy-paste error or something that needs explanation. As written, it undercuts the compositional story. The URPC2020 section does not say whether the model was fine-tuned or zero-shot, nor how the class head was adapted, so the 87.2% claim is not reproducible from the text. Also, the YOLOv10 and YOLOv11 references look wrong (YOLOv10 is not arXiv:2305.10530, and the YOLOv11 TPAMI citation smells invented). No code or dataset is released.\n\nNone of these are fatal to the underlying idea, but they are exactly the things that need fixing before the numbers can be trusted.\n\nWho it is for: people working on lightweight detectors for agriculture and aquaculture, and referees who value dataset contributions in applied CV. I would send it to peer review, because the dataset has value and the approach is sane, but I would ask for a corrected ablation, multi-run statistics, and a clear URPC protocol before accepting.\n\nRecommendation: engage with it, but make the experimental transparency a condition of acceptance.","headline":"A routine YOLOv8n tweak with a genuinely new shrimp-disease dataset, but the reported gains sit on a thin single-run evaluation and one implausible ablation row.","tokens_in":9738,"tokens_out":2322,"would_cite":false,"duration_ms":24827,"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 YOLOv8n variant with a reparameterized detection head, a multi-scale EMCM module, and an improved attention mechanism reports 92.7% mAP@0.5 on a three-class shrimp disease dataset while cutting parameters by 32.3%.","keywords":["shrimp disease detection","YOLOv8n","lightweight object detection","reparameterized detection head","multi-scale convolution","attention mechanism","URPC2020","mAP@0.5"],"falsifier":"Re-train the baseline and the full model on the same 1,047 images under several random 9:1 splits (or on an independent public shrimp-disease dataset) and compare the mAP differences; if the 92.7% versus 89.7% gap shrinks below the run-to-run standard deviation, the claimed improvement is not established.","tokens_in":8742,"feed_emoji":"🦐","tokens_out":6124,"duration_ms":61285,"temperature":0.7,"pith_summary":"This paper sets out to show that a standard YOLOv8n detector can be made simultaneously lighter and more accurate for shrimp disease detection. The proposed network swaps in a reparameterized lightweight detection head (RLDD), replaces standard bottleneck convolutions with a multi-scale EMCM module, and adds an improved SegNext-style attention mechanism. On a self-collected dataset of 1,047 images covering white spot, black gill, and yellow gill disease, the model reports 92.7% mAP@0.5 with 2.1 million parameters — a 3 percentage point gain over YOLOv8n at a 32.3% parameter reduction. If the result holds, it means accurate disease screening could run on low-cost, resource-constrained devices in shrimp farms rather than requiring cloud servers.","feed_headline":"Shrimp-disease model: 92.7% mAP, 32% fewer params than YOLOv8n","feed_subtitle":"A 2.1M-parameter YOLOv8n variant for spotting white spot, black gill, and yellow gill disease on low-cost hardware.","key_machinery":"The load-bearing mechanism is the combination of three architectural substitutions. RLDD uses reparameterized convolutions: during training the detection head has 3×3, 1×1, and 3×3 average-pool branches; at inference these merge losslessly into one 3×3 convolution, cutting parameters while keeping multi-scale capacity. EMCM splits input channels, passes one branch straight through as original features and the other through parallel 3×3 and 5×5 convolutions, then concatenates and compresses with a 1×1 convolution; replacing the standard Bottleneck in selected C2f blocks gives multi-scale receptive fields. The SegNext_Attention module builds an attention map from depthwise and multi-scale depthwise convolutions and multiplies it onto the feature map, letting the model focus on disease-relevant regions.","core_discovery":"On its own terms, the paper claims that a YOLOv8n detector can be made lighter and more accurate for shrimp disease detection by stacking three changes: an RLDD head that trains as a multi-branch structure and collapses to a single 3×3 convolution at inference; a C2f-EMCM module that splits channels and extracts 3×3 and 5×5 multi-scale features; and an improved SegNext_Attention mechanism that computes pixel-level attention to suppress background. The full system reports 92.7% mAP@0.5, 87.5% precision, and 88.9% recall on the self-built dataset with 2.1M parameters, versus 89.7% mAP and 3.1M parameters for YOLOv8n; on URPC2020 it reports 87.2% mAP against 83.1% for YOLOv8n. The paper concludes that the architecture achieves the desired accuracy-efficiency balance for shrimp aquaculture monitoring.","pith_inferences":["The paper's evidence would be stronger with repeated runs; absent that, the practical takeaway should be treated as an existence proof that the three-module stack can work, not as a guaranteed improvement on a new farm's dataset.","Because the EMCM+SegNext ablation lands exactly at baseline mAP while the full stack improves, the modules appear to interact rather than simply add; testing each on more varied data could clarify where the synergy comes from.","The URPC2020 result would be more compelling if the protocol (zero-shot versus fine-tuned) were explicit, since zero-shot generalization across domains would be the stronger claim."],"forward_implications":["The model can be deployed on edge devices for near-real-time shrimp disease screening: 2.1M parameters and a 32.3% parameter cut relative to YOLOv8n at equal or better mAP.","The RLDD reparameterized head alone yields a 25.8% parameter reduction with a precision gain, so reparameterization of the detection head is a generally applicable lightweighting technique.","The architecture transfers to a different underwater detection domain (URPC2020) with a 4.1-point mAP gain over YOLOv8n, suggesting the module combination generalizes beyond shrimp disease."],"supporting_citations":[{"why":"Defines the YOLO detection paradigm that frames the paper's end-to-end regression approach.","marker":"[3]"},{"why":"Supplies the SegNext_Attention/MSCA module that the paper adapts for feature enhancement.","marker":"[13]"},{"why":"Provides the YOLOv5 baseline used in the comparison experiments.","marker":"[14]"},{"why":"Provides the YOLOv8n baseline that the paper modifies and compares against.","marker":"[15]"},{"why":"Provides the YOLOv10 lightweight baseline in the comparison table.","marker":"[16]"},{"why":"Provides the YOLOv11 baseline in the comparison table.","marker":"[17]"},{"why":"Provides the FasterNet baseline used as a lightweight efficiency reference.","marker":"[18]"},{"why":"Provides the BiFPN baseline used as a feature-fusion comparison.","marker":"[19]"}],"fun_headline_variants":["Shrimp disease spotter: lighter and more accurate than YOLOv8n","YOLOv8n slimmed for shrimp disease: 92.7% mAP, 32% fewer params","Lightweight YOLOv8n catches shrimp disease with 92.7% mAP","2.1M-param YOLOv8n detects shrimp disease at 92.7% mAP","Shrimp disease AI: 32% lighter YOLOv8n, 3% better mAP"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported improvements depend on the single 9:1 split of a private 1,047-image dataset, with no held-out test set and no repeated runs, so the 3-point mAP gain could change if the data or split changes.","fun_headline_variants_meta":{"raw":{"variants":["Shrimp disease spotter: lighter and more accurate than YOLOv8n","YOLOv8n slimmed for shrimp disease: 92.7% mAP, 32% fewer params","Lightweight YOLOv8n catches shrimp disease with 92.7% mAP","2.1M-param YOLOv8n detects shrimp disease at 92.7% mAP","Shrimp disease AI: 32% lighter YOLOv8n, 3% better mAP"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000656,"raw_usage":{"total_tokens":3045,"prompt_tokens":1030,"completion_tokens":2015,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":646,"completion_tokens_details":{"reasoning_tokens":1888}},"tokens_in":646,"tokens_out":2015,"duration_ms":15173,"temperature":1.0,"reasoning_tokens":1888,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:31:17.857778+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-train the baseline and the full model on the same 1,047 images under several random 9:1 splits (or on an independent public shrimp-disease dataset) and compare the mAP differences; if the 92.7% versus 89.7% gap shrinks below the run-to-run standard deviation, the claimed improvement is not established.","supporting_citations":[{"cited_title":"Apple ste m/calyx real -time recognition using YOLO -v5 algo rithm for fruit automatic loading syst em[J]","cited_arxiv_id":null,"evidence_quote":"Provides the YOLOv5 baseline used in the comparison experiments."},{"cited_title":"Re search on traffic sign detection based on improve d YOLOv8[J]","cited_arxiv_id":null,"evidence_quote":"Provides the YOLOv8n baseline that the paper modifies and compares against."},{"cited_title":"Personalized action suggestions in low-code automation platforms","cited_arxiv_id":"2305.10530","evidence_quote":"Provides the YOLOv10 lightweight baseline in the comparison table."},{"cited_title":"YOLOv11: Dynami c feature fusion for real -time detection[J]","cited_arxiv_id":null,"evidence_quote":"Provides the YOLOv11 baseline in the comparison table."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the FasterNet baseline used as a lightweight efficiency reference."},{"cited_title":"EfficientDet: Sc alable and efficient object detection[C]//Proceeding s of 2020 IEEE/CVF Conference on Computer Vi sion and Pattern Recognition (CVPR)","cited_arxiv_id":null,"evidence_quote":"Provides the BiFPN baseline used as a feature-fusion comparison."}],"review_version":1}