{"id":"cef2b3d9-5c6b-4174-86fb-71e2b20e19f7","arxiv_id":"2506.21135","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A YOLOv5 variant with BiFPN, directional detail enhancement, and two attention fusion modules reports state-of-the-art mAP on GC10-DET and DAGM2007.","lead":"The paper stacks known components (BiFPN, asymmetric convolution, and attention-weighted fusion) onto YOLOv5 and reports higher mAP on two industrial defect benchmarks. It is a serviceable engineering recipe, but the gains are modest and the evaluation has uncontrolled baselines and no error bars.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Claimed SOTA-level accuracy and robustness rests on copied numbers from 200/500-epoch protocols and on no explicit robustness evaluation; the only same-protocol baseline is YOLOv5s, and the DDFM module alone reduces mAP.","rationale":"The reader's weakest assumption correctly identifies the biggest accuracy-related weakness: most SOTA numbers are copied from other papers with different training budgets, so the reported margin may partly reflect protocol rather than architecture. I agree this is load-bearing. But the central claim has a second, more under-supported component: robustness. The paper provides no perturbation-based robustness evaluation at all, despite the abstract and conclusion claiming robustness. I also note that the paper openly discloses the protocol mismatch, which is a fair limitation rather than a hidden flaw. The DDFM ablation being negative without the attention modules is additional evidence that the named contribution is not independently validated. These issues together keep the central claim conditional: the improvement over the reproduced YOLOv5s baseline is meaningful, and the architecture may well work, but the broader SOTA and robustness statements require matched-protocol reruns, error bars, and robustness stress tests. No change to the reader's CONDITIONAL verdict is needed.","tokens_in":11283,"tokens_out":3765,"duration_ms":44662,"concrete_test":"Rerun GC10-DET under the paper's exact protocol (8:1:1 split, 640x640 input, 250 epochs, batch size 8, SGD lr 0.01) for YOLOv7, YOLOv9s, DSL-YOLO, YOLOv5s-improved, and YOLOv5s, with at least 3 seeds each, and compare the mAP50 distributions; if any baseline reaches 74.9% or comes within 1 point, the claimed SOTA margin weakens. Separately, take the trained YOLO-FDA and YOLOv5s checkpoints and evaluate under Gaussian noise, blur, and 0.5x-2x scale perturbations on GC10-DET; if mAP50 degrades similarly for both models, the robustness claim is unsupported by the current experiments.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract claims YOLO-FDA consistently outperforms existing state-of-the-art methods in both accuracy and robustness. To support that claim, the paper needs (a) SOTA comparisons under matched training conditions and (b) an operational measure of robustness. Neither is established. In Table 1, all non-baseline comparators (YOLOv7, YOLOv9s, DSL-YOLO, etc.) are quoted from papers with 200 or 500 training epochs, batch size 32, and different GPUs, versus the authors' 250 epochs and batch size 8; the paper itself acknowledges this in the text. The 72.5-to-74.9 mAP50 gap over DSL-YOLO is therefore confounded by training protocol. On DAGM2007, Table 2 claims some same-environment baselines, but the gain over YOLOv9 is only 0.8 mAP50-95 and no error bars are reported, so consistent outperformance is not established with any variance estimate. Meanwhile, the ablation in Fig. 5 shows BiFPN+DDFM is worse than BiFPN alone, meaning the sole detail-enhancement module contributes only in combination with AC/CAF, and the reduction ratio r in Eqs. 4 and 9 is never specified, blocking exact reproduction. Finally, no experiment perturbs scale, noise, lighting, or texture, yet the abstract explicitly claims robustness. No manuscript passage supplies the missing robustness evidence.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes YOLO-FDA, a YOLOv5-based detector for surface defect detection that replaces the PANet neck with a BiFPN-style architecture, adds a Detail-directional Fusion Module (DDFM) using asymmetric convolutions on a low-level feature path, and introduces two attention-based fusion mechanisms (AC and CAF) at three-path fusion nodes. Experiments are reported on GC10-DET and DAGM2007. The authors claim consistent state-of-the-art accuracy and robustness, with the main quantitative results being 74.9% mAP50 on GC10-DET (4.6 points over their own YOLOv5s baseline) and 67.9% mAP50-95 on DAGM2007 (2.7 points over YOLOv5s). Ablation studies on GC10-DET show that BiFPN helps, the addition of DDFM alone slightly hurts, and the AC/CAF modules improve results, with the full model giving the best performance.","tokens_in":11558,"tokens_out":2406,"duration_ms":27246,"significance":"If the empirical claims are taken at face value, the work is an incremental but potentially useful engineering contribution to YOLO-based surface defect detection. The controlled baseline improvement of 4.6 mAP50 over a re-run YOLOv5s is a genuine positive result, and the combination of directional detail enhancement with attention-weighted fusion is a reasonable design direction. However, the headline claim of consistent superiority over prior state-of-the-art is not supported by the evidence as presented: almost all SOTA comparisons are quoted from papers with different training protocols, and the robustness claim has no dedicated experiments. The novelty is moderate, and the paper does not provide code, model weights, or a fully specified architecture, which limits reproducibility and immediate practical uptake.","major_comments":[{"comment":"The central claim of consistent outperformance over existing state-of-the-art rests on comparisons that are not matched in training protocol. The paper itself acknowledges that YOLOv3, YOLOv7tiny, YOLOv5n, YOLOv8n, YOLOv7, and DSL-YOLO are taken from related work [40] with 200 training epochs and a different GPU, and that YOLOv9s and YOLOv5s-improved come from [41] with 500 epochs and batch size 32, whereas the authors use 250 epochs and batch size 8. The margin over DSL-YOLO (72.5 vs. 74.9 mAP50) could plausibly shrink or reverse under identical training budgets. To support the SOTA claim, the authors should re-run the key competitors under their own protocol or, failing that, weaken the claim to 'improvement over the YOLOv5s baseline' and clearly label quoted numbers as incomparable.","section":"§4.4, Table 1"},{"comment":"The abstract and conclusion claim robustness, but no experiment in the paper perturbs scale, noise, lighting, texture, or any other input condition. The visualization section shows only a few qualitative examples. The term 'robustness' is used without an operational definition or quantitative evaluation. The authors should either add a robustness evaluation (e.g., noise injection, scaling, or contrast perturbations) or remove the robustness claim from the abstract and conclusions.","section":"Abstract and §4.4"},{"comment":"The ablation shows that adding DDFM to BiFPN decreases mAP50 compared with BiFPN alone, so the sole detail-enhancement module contributes positively only when combined with the AC/CAF attention modules. The paper's explanation that simple concatenation causes redundancy is speculative and not supported by the ablation design. This is a load-bearing point because DDFM is one of the paper's three named contributions. The authors should provide a more thorough analysis, e.g., ablating DDFM with each attention module separately, or reporting intermediate feature statistics, to clarify whether DDFM has any independent value.","section":"§4.5, Fig. 5"},{"comment":"The reduction ratio r that defines the bottleneck channels C' = C/r in the AC module (Eq. 4) and the CAF module (Eq. 9) is never given a value. Without this hyperparameter, the architecture is not fully specified and the experiments cannot be exactly reproduced. The paper should state the value of r used in all experiments (and, if varied, report sensitivity).","section":"Eqs. (4) and (9), §4.3"},{"comment":"No variance estimates are reported for any of the runs, and each configuration appears to be trained once. On DAGM2007, the gain over the same-environment YOLOv9 baseline is only 0.8 mAP50-95 (67.9 vs. 67.1), which is likely within the run-to-run variation of a single-seed training. The authors should report results over multiple seeds (at least 3) with mean and standard deviation for at least their own model and the same-environment baselines, so the reader can judge whether the improvements are statistically meaningful.","section":"Tables 1 and 2"}],"minor_comments":[{"comment":"The evaluation metrics sentence says 'the Precision, Recall, and mAP are used to quantitatively evaluate the detection performance on the GC10-DET and DAGM2007, respectively,' which is grammatically unclear; it should specify that mAP50 is used for GC10-DET and mAP50-95 for DAGM2007 in the corresponding tables.","section":"§4.2"},{"comment":"The figure caption simply says 'Examples of surface defect from DAGM2007 and GC10-DET,' but the text refers to colored boxes in the second picture; the figure should label the defect categories (inclusion, weld, etc.) directly or the caption should explain the colored boxes.","section":"Fig. 1"},{"comment":"There are several typographical and formatting issues, including 'BiPFN' for BiFPN in the ablation discussion, 'indicator' used loosely for 'metric,' and incomplete reference entries (e.g., references [22], [33], [35] have no full bibliographic details). These should be cleaned up.","section":"Throughout"},{"comment":"The conclusion repeats the claim of 'multiscale robustness' without qualifications; given the absence of robustness experiments, this should be rephrased to reflect what is actually demonstrated.","section":"§5, Conclusion"}],"recommendation":"major_revision","confidential_remarks":"The paper's self-citation pattern (several references to F. Shen et al. on virtual try-on and human-centric generation) seems tangential to the main topic and may be an attempt to inflate citation counts; the editor may want to check whether these citations are necessary. The manuscript is also very short for a full journal/conference paper and resembles an extended abstract; the authors should be encouraged to add more experimental detail and analysis. The core idea is not flawed, but the evidence for the SOTA claim needs substantial strengthening."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a modest, routine YOLOv5 variant that reports a genuine gain over its own baseline but does not support the abstract's claims about state-of-the-art accuracy and robustness. The paper is transparent about its comparison protocol, which makes it fixable rather than misleading.\n\nWhat's new: replacing PANet with BiFPN, adding asymmetric convolutions for directional detail, and two attention weighting schemes (AC and CAF). None of these techniques are new, but the combination on GC10-DET and DAGM2007 yields +4.6 mAP50 and +2.7 mAP50-95 over a same-setting YOLOv5s re-run. That controlled baseline is the paper's strongest evidence. The ablation is reasonably complete, and the authors disclose that DDFM alone hurts performance; the explanation (concatenation causes redundancy) is post hoc but plausible.\n\nSoft spots, in order of severity. First, all SOTA comparisons in Table 1 are quoted from other papers with 200 or 500 training epochs, batch size 32, and different GPUs, versus the authors' 250 epochs and batch size 8. The paper itself admits this, so the \"consistently outperforms\" claim is not supported. Second, no repeated runs or error bars anywhere, so we have no variance estimate for the reported margins. Third, the abstract claims robustness, but no experiment perturbs scale, noise, lighting, or texture. That claim should be either tested or removed. Fourth, the reduction ratio r in Eqs. (4) and (9) is never specified, blocking exact reproduction. Fifth, the related work cites several image generation papers (ImagGarment, ImagDressing) with no evident connection to defect detection; that looks like padding.\n\nThe math is standard, the writing is clear, and the authors are honest about protocol mismatches. This is exactly the kind of paper that benefits from peer review: the central claim is testable, the weaknesses are identifiable, and a referee can insist on matched baselines, multiple seeds, and a robustness experiment. I would not desk-reject it, but I would send it to review expecting major revisions. I would not cite it in my own work.\n\nRecommendation: accept for peer review, with instructions to focus on baseline matching and variance; expect a conditional accept at best.","headline":"Routine YOLOv5 variant with a real controlled-baseline gain, but SOTA and robustness claims outrun the evidence.","tokens_in":12018,"tokens_out":2072,"would_cite":false,"duration_ms":24066,"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":"The paper claims that YOLO-FDA, a YOLOv5 variant with a BiFPN neck, directional-detail fusion, and two attention fusion modules, improves surface-defect detection accuracy on steel and synthetic-texture benchmarks, reporting 74.9% mAP50…","keywords":["surface defect detection","YOLOv5","BiFPN","asymmetric convolution","attention fusion","GC10-DET","DAGM2007"],"falsifier":"Re-run every compared detector, including DSL-YOLO, YOLOv9s, and the multi-layer fusion variant, on GC10-DET and DAGM2007 with exactly the paper's settings of 250 epochs, batch size 8, and the same data split, and check whether YOLO-FDA still leads in mAP50 and mAP50-95; if the margin shrinks to noise or reverses, the claimed state-of-the-art advantage does not hold. A second check is whether removing DDFM while keeping AC and CAF changes the score, since the paper's own ablation shows DDFM alone lowers mAP.","tokens_in":11085,"feed_emoji":"🔍","tokens_out":4708,"duration_ms":48765,"temperature":0.7,"pith_summary":"This paper targets industrial surface defect detection, where defects are small, elongated, and embedded in textured backgrounds. It claims that standard YOLO-style feature fusion drops the fine directional detail these defects need, and that a modified YOLOv5, named YOLO-FDA, fixes this by combining a BiFPN neck, a Detail-directional Fusion Module that uses asymmetric 1x7 and 7x1 convolutions, and two attention-based fusion strategies. On the GC10-DET steel dataset it reports 74.9% mAP50, 4.6 points above its own re-run YOLOv5s baseline, and on DAGM2007 it reports 67.9% mAP50-95, 2.7 points above the baseline. The paper's conclusion is that this combination improves both accuracy and robustness across defect types and scales.","feed_headline":"YOLO-FDA lifts steel defect detection by 4.6 mAP points","feed_subtitle":"Attention-weighted fusion plus asymmetric convolution catches small, elongated surface flaws that YOLOv5 misses.","key_machinery":"The central mechanism is the network-neck redesign: a BiFPN-style bidirectional multi-scale fusion; the Detail-directional Fusion Module (DDFM), which fuses the second-lowest and fourth backbone layers and applies 1x7 and 7x1 asymmetric convolutions to expose directional texture; and two learnable fusion rules at the three-path merge nodes, namely Attention-weighted Concatenation (AC), which weights each feature map before concatenation, and Cross-layer Attention Fusion (CAF), which computes a softmax-weighted sum across layers. Together these modules carry the claim that fine detail can be preserved early and redundant cross-scale information can be suppressed late.","core_discovery":"YOLO-FDA claims that replacing YOLOv5's PANet neck with BiFPN and adding targeted detail and attention modules yields consistent gains on surface-defect benchmarks. The Detail-directional Fusion Module enriches the second-lowest feature layer by upsampling the fourth backbone layer, concatenating it with low-level features, and adding two asymmetric convolutions (1x7 horizontal, 7x1 vertical) so elongated and directional defects are not lost. Attention-weighted Concatenation scales each input feature map before concatenation to preserve all channels, while Cross-layer Attention Fusion learns a softmax-weighted sum across layers to suppress redundancy. The reported results are 74.9% mAP50 on GC10-DET and 67.9% mAP50-95 on DAGM2007, with ablation experiments showing the full combination outperforms any single module.","pith_inferences":["The paper's ablation shows DDFM by itself reduces mAP; an editorial inference is that the net gain is carried by AC and CAF, so a simpler model with only attention fusion might match YOLO-FDA on these benchmarks.","Because most comparison numbers come from other papers with different training epochs, batch sizes, and GPUs, the true rank against DSL-YOLO and YOLOv9s may change under uniform training conditions.","The same two-stage fusion recipe could transfer to other one-stage detectors handling elongated objects, such as PCB defects or fabric flaw detection, though this is beyond the paper's evidence."],"forward_implications":["If the reported gains hold, industrial inspection systems can adopt YOLO-FDA directly on top of existing YOLOv5 deployments for a 4.6-point mAP50 gain on steel surfaces without changing the detection head.","The AC and CAF fusion rules imply that preserving channel diversity early and mixing additively late improves small-defect accuracy in feature-concatenation networks.","Asymmetric 1x7 and 7x1 convolutions provide a concrete recipe for detecting defects with high aspect ratio, such as welds and inclusions.","On DAGM2007, the C1 category gain (84.5% versus 75.9% for YOLOv5s) suggests the module combination particularly helps defect classes with lower baseline detectability.","The consistent gains on two different benchmarks indicate the neck-level changes are portable across data domains rather than tuned to one texture type."],"supporting_citations":[{"why":"Supplies the GC10-DET steel defect dataset on which the main mAP50 comparison is run.","marker":"[33]"},{"why":"Supplies the DAGM2007 dataset and its weak labels, which the paper converts to bounding boxes for the second benchmark.","marker":"[34]"},{"why":"Defines the YOLOv5 baseline and architecture that YOLO-FDA modifies.","marker":"[22]"},{"why":"Source of the DSL-YOLO and several YOLO comparison numbers on GC10-DET.","marker":"[40]"},{"why":"Source of the YOLOv9s and YOLOv5s-improved comparison numbers, trained with 500 epochs and batch size 32.","marker":"[41]"},{"why":"Provides the HIC-YOLOv5 small-object comparison result on DAGM2007.","marker":"[46]"}],"fun_headline_variants":["YOLO-FDA uses asymmetric convolutions to catch thin surface defects","Attention fusion boosts YOLOv5 for surface defect detection","New YOLO-FDA module finds small defects with directional convolutions","BiFPN and attention cut feature noise in defect detection","Detail-enhanced YOLO-FDA improves surface defect spotting"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that published state-of-the-art numbers trained with different epochs, batch sizes, and GPUs can be compared directly against the paper's own runs, since only the YOLOv5s baseline was trained under identical conditions.","fun_headline_variants_meta":{"raw":{"variants":["YOLO-FDA uses asymmetric convolutions to catch thin surface defects","Attention fusion boosts YOLOv5 for surface defect detection","New YOLO-FDA module finds small defects with directional convolutions","BiFPN and attention cut feature noise in defect detection","Detail-enhanced YOLO-FDA improves surface defect spotting"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000416,"raw_usage":{"total_tokens":2148,"prompt_tokens":946,"completion_tokens":1202,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":562,"completion_tokens_details":{"reasoning_tokens":1115}},"tokens_in":562,"tokens_out":1202,"duration_ms":11660,"temperature":1.0,"reasoning_tokens":1115,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:31:37.666760+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run every compared detector, including DSL-YOLO, YOLOv9s, and the multi-layer fusion variant, on GC10-DET and DAGM2007 with exactly the paper's settings of 250 epochs, batch size 8, and the same data split, and check whether YOLO-FDA still leads in mAP50 and mAP50-95; if the margin shrinks to noise or reverses, the claimed state-of-the-art advantage does not hold. A second check is whether removing DDFM while keeping AC and CAF changes the score, since the paper's own ablation shows DDFM alone lowers mAP.","supporting_citations":[{"cited_title":"dataset., in : https://github.com/lvxiaoming2019/GC10-DET-metallic-surface-defect- datasets","cited_arxiv_id":null,"evidence_quote":"Supplies the GC10-DET steel defect dataset on which the main mAP50 comparison is run."},{"cited_title":"Weakly supervised learning of a classifier for unusual event detection,","cited_arxiv_id":null,"evidence_quote":"Supplies the DAGM2007 dataset and its weak labels, which the paper converts to bounding boxes for the second benchmark."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the YOLOv5 baseline and architecture that YOLO-FDA modifies."},{"cited_title":"Research on a metal surface defect detection algorithm based on dsl-yolo,","cited_arxiv_id":null,"evidence_quote":"Source of the DSL-YOLO and several YOLO comparison numbers on GC10-DET."},{"cited_title":"Steel surface defect detection based on multi-layer fusion networks,","cited_arxiv_id":null,"evidence_quote":"Source of the YOLOv9s and YOLOv5s-improved comparison numbers, trained with 500 epochs and batch size 32."},{"cited_title":"Hic-yolov5: Improved yolov5 for small object detection,","cited_arxiv_id":null,"evidence_quote":"Provides the HIC-YOLOv5 small-object comparison result on DAGM2007."}],"review_version":1}