{"id":"f6efc514-f06f-47d9-b574-3304e875b8be","arxiv_id":"2507.14010","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A two-stage deep learning pipeline (DenseNet-169 then DeepLabV3+) for tunnel crack classification and segmentation achieves 92.23% classification accuracy and 57.01% segmentation IoU on the NUAACrack-2000 dataset.","lead":"This paper combines two standard deep learning tools, DenseNet-169 for image classification and DeepLabV3+ for segmentation, to detect cracks in tunnel lining images, and applies Score-CAM to visualize what the segmentation model focuses on. A smart generalist might read it to see an engineering pipeline for automating tunnel inspection, though the method is largely a practical application of existing techniques.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed superiority of the two-stage DenseNet-169/DeepLabV3+ pipeline is not established: Table 2 comparators collapse to the majority-class baseline, Table 3 IoU margins are within noise, and no end-to-end result is reported.","rationale":"Read in good faith, this is an engineering evaluation rather than a novel architecture proposal. The central claim requires two conditions: the two-stage design preserves or improves detection relative to single-stage segmentation, and the chosen models genuinely beat the alternatives. The first condition is untested because the classification stage discards about 8.2% of crack images before segmentation, and no end-to-end metric is provided; the likely product-level recall is below the segmentation-only recall, which directly undermines the claimed accuracy advantage of combining the two stages. The second condition is weakened by internal evidence: three classification comparators land exactly on the majority-class baseline, indicating training collapse rather than architectural inferiority, and the segmentation IoU advantage over UNet is only 0.17 percentage points without error bars. These are internal-evidence problems, not disagreements with field consensus, and they are directly testable. The paper does have independent support: it uses a public dataset, provides qualitative visual explanations, and reports a clear experimental setup. The absence of code and the lack of repeated runs are additional limitations but not the central flaw. If the proposed tests show that properly trained comparators still lose by meaningful margins and the end-to-end pipeline improves recall, the paper's claims stand; otherwise the conclusions should be softened to a feasibility demonstration rather than demonstrated superiority.","tokens_in":10230,"tokens_out":4840,"duration_ms":60038,"concrete_test":"Retrain the six Table 2 comparators with class-balanced sampling and the same 100-epoch budget; verify that their reported accuracies exceed the 75.65% majority baseline and report confusion matrices. For Table 3, run DeepLabV3+ and UNet from five random seeds and report mean +/- std IoU and F1 on the held-out 146 images; if the gap is less than one standard deviation, the outperformance claim is unsupported. Finally, chain the saved stage-1 classifier and stage-2 segmenter over the full 193-image test set and compute end-to-end crack-pixel recall and precision to quantify the actual benefit of the two-stage pipeline.","verdict_should_be":"UNCHANGED","load_bearing_attack":"For the central claim to hold, the comparisons must be fair and statistically meaningful, and the two-stage design must be shown to improve detection. None of these conditions is currently met. In Table 2, EfficientNet-B0, Swin Transformer, and Vision Transformer all achieve exactly 75.65% accuracy, which equals the majority-class baseline (146 crack images out of 193 test images). This indicates these comparators collapsed to predicting the majority class, so the reported 16.58% improvements over them do not demonstrate superiority over properly trained alternatives. In Table 3, DeepLabV3+ achieves 57.01% IoU versus 56.84% for UNet and 56.42% for UNet++; with a 146-image test set and no repeated runs or confidence intervals, differences of 0.17 and 0.59 percentage points are within run-to-run noise. The paper even reports a lower F1 score than UNet (67.44 vs. 67.68) and lower recall than UNet (82.12 vs. 83.91). Separately, the central two-stage benefit is asserted but never measured: the classification stage misses 12 of 146 crack images (about 8.2% false negatives), and those images never reach the segmentation stage. No end-to-end detection metric is reported, so the combined recall and precision of the pipeline are unknown; a simple product estimate would put crack-pixel recall near 75%, below the 82.12% of the segmentation model alone. The claims that the two-stage framework outperforms other models and improves detection accuracy therefore rest on degenerate baselines, statistically fragile margins, and an unmeasured end-to-end advantage.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-stage deep-learning framework for tunnel lining crack detection: DenseNet-169 first classifies each image as crack or defect-free, and DeepLabV3+ then segments cracks in the images classified as containing cracks. Score-CAM is applied to the segmentation model to produce visual explanations. Experiments on the open NUAACrack-2000 dataset report 92.23% classification accuracy and 39.80 FPS for DenseNet-169, and 57.01% IoU / 67.44% F1 for DeepLabV3+. Comparisons are made with six classification models and four segmentation models, and the paper claims that the proposed two-stage method outperforms these alternatives.","tokens_in":10521,"tokens_out":3423,"duration_ms":42870,"significance":"If the claims were established, the work would be of practical interest for tunnel inspection, because it combines image-level screening with pixel-level crack segmentation and adds interpretability through Score-CAM. The manuscript is clearly organized and uses a publicly available dataset, and it compares against a reasonable set of CNN and Transformer baselines. The visual explanations in Section 4.4 provide a useful qualitative check on the segmentation model's behavior. However, as presented, the central claims are not supported by the evidence: key classification baselines collapse to the majority-class baseline, the segmentation margins over UNet and UNet++ are within run-to-run noise, and no end-to-end metric for the two-stage pipeline is reported. The paper's empirical contribution is therefore currently uncertain until these issues are addressed.","major_comments":[{"comment":"The claimed classification superiority is not established because several comparators collapse to the majority-class baseline. EfficientNet-B0, Swin Transformer, and Vision Transformer all report 75.65% accuracy, which is exactly the accuracy of always predicting the crack class on the 193-image test set (146/193). The stated improvements of 16.58% over these models therefore compare DenseNet-169 against effectively untrained or degenerate classifiers. The authors should re-tune or re-train these baselines, verify that their training protocols lead to non-degenerate predictions (e.g., report confusion matrices or a majority-class baseline), and then reassess the claimed improvement.","section":"§4.2.3, Table 2"},{"comment":"The segmentation comparison does not support the claim that DeepLabV3+ outperforms other state-of-the-art models. DeepLabV3+ achieves 57.01% IoU versus 56.84% for UNet and 56.42% for UNet++, differences of 0.17 and 0.59 percentage points on a 146-image test set. The paper reports no repeated runs, confidence intervals, or statistical significance tests, and DeepLabV3+ actually has a lower F1 score (67.44 vs. 67.68) and lower recall (82.12 vs. 83.91) than UNet. These margins are within plausible run-to-run variation, so the reported numbers do not demonstrate a reliable advantage. The authors should provide uncertainty estimates, e.g., via multiple training runs with different seeds or a paired bootstrap test on the test set.","section":"§4.3.3, Table 3"},{"comment":"The central two-stage benefit is asserted but never measured end-to-end. The classification stage misclassifies 12 of 146 crack images in the test set, and those images are then excluded from the segmentation evaluation; no metric reports what the full pipeline produces on the entire 193-image test set. The paper also does not compare the end-to-end speed of the two-stage pipeline against a one-stage segmentation model on all images. To support the claim that the two-stage design improves detection accuracy and efficiency, the authors should report an end-to-end measure (e.g., pipeline recall/precision over all test images, combined with a latency/throughput comparison of the whole system).","section":"§2.1, §4.3.2"}],"minor_comments":[{"comment":"The equations and surrounding notation are poorly typeset: e.g., \"k lA\", \"H s Up A\", \"co c S re CAML\", and the use of roman and italic letters is inconsistent. The authors should rewrite the Score-CAM equations cleanly, defining all symbols (including X_b, f, and the upsampling operator).","section":"§2.4, Eqs. (1)–(4)"},{"comment":"The abstract claims that the segmentation model's F1 score \"outperforms other state-of-the-art models,\" but Table 3 shows UNet has a higher F1 score (67.68 vs. 67.44). The abstract should be corrected to state the actual pattern of results.","section":"Abstract and §4.3.3"},{"comment":"Section 3 states that the original images have a pixel size of 512×375, while Section 4.3.2 states that images were resized to 512×384 before segmentation. This discrepancy, and the reason for the different aspect ratio, should be clarified.","section":"§3 and §4.3.2"},{"comment":"Several references are incomplete or inconsistently formatted. For instance, the Zhou et al. (2023) entry reads \"Computer aided Civil Eng mice.13003\" instead of a full citation, and the two Feng et al. (2023) entries are not distinguished with year suffixes.","section":"References"},{"comment":"The text states that Swin Transformer and Vision Transformer \"wrongly classified all defect-free images as crack images,\" but no confusion matrix is shown for any model. Reporting confusion matrices for all classification models would make the comparison transparent and would clarify the majority-class collapse.","section":"§4.2.3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript addresses an applied problem well within the journal's scope, and the dataset choice and explanatory-evaluation component are appropriate. The main concern is evidentiary: the classification baselines are degenerate, the segmentation margins are within noise, and the two-stage pipeline is not evaluated as a whole. I recommend major revision, not rejection, because these issues can in principle be fixed by additional experiments (re-training baselines, adding error bars/significance tests, and reporting end-to-end pipeline metrics). If the authors cannot supply such evidence, the performance claims should be substantially downgraded to a qualitative demonstration."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a clean but entirely conventional engineering application of DenseNet-169 for classification and DeepLabV3+ for segmentation on the NUAACrack-2000 tunnel dataset. The abstract's numbers are real, but the paper's central claim—that the two-stage framework outperforms other models and improves detection—is much weaker than the evidence supports.\n\nWhat's new: the specific DenseNet-169 + DeepLabV3+ + Score-CAM combination has not been reported on this dataset, and the Score-CAM heatmaps inside the DeepLabV3+ decoder are a small useful addition. The model descriptions and Score-CAM equations are correct, and the experimental setup is clearly documented.\n\nThe soft spots are exactly where the stress-test note lands. In Table 2, EfficientNet-B0, Swin Transformer, and Vision Transformer all hit 75.65% accuracy, which equals the majority-class baseline (146 of 193 test images are cracks). Those comparators collapsed to predicting the majority class, so the reported 16.58% improvements over them prove nothing about DenseNet-169's superiority. The authors even note that Swin and ViT misclassified every defect-free image, confirming these baselines were not properly trained or tuned. A fair comparison would need hyperparameter tuning or class weighting for the comparators. In Table 3, the IoU margins over UNet (0.17 pp) and UNet++ (0.59 pp) on a 146-image test set are within run-to-run noise, and the paper itself reports lower F1 than UNet and lower recall than both UNet and UNet++. No error bars, no repeated runs, no significance tests.\n\nThe bigger structural flaw is that the two-stage benefit is never measured. The classification stage misses 12 of 146 crack images (8.2% false negatives), and those images never reach segmentation. No end-to-end metric is reported. A simple product estimate puts crack-pixel recall near 75% (0.9223 × 0.8212), below the 82.12% of the segmentation model run alone. So the claimed efficiency and accuracy advantage of the two-stage design is asserted, not demonstrated.\n\nWhat I do credit: the paper acknowledges its own limitations (small dataset, class imbalance, need for preprocessing and lightweight models), and the visual explanation section is honest interpretation work. It is a plausible engineering report, not a research advance.\n\nBottom line: this deserves a serious referee only if the editor insists on a revised version with fixed baselines, error bars, and a direct end-to-end evaluation. As it stands, I would not cite the quantitative claims in my own work, though the paper could become a useful data point after major revision.","headline":"Competent but very conventional two-stage crack detection paper whose 'superiority' claims rest on degenerate classification baselines, within-noise segmentation margins, and an unmeasured end-to-end benefit.","tokens_in":11108,"tokens_out":1938,"would_cite":false,"duration_ms":24426,"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 two-stage deep learning pipeline, using DenseNet-169 for classification and DeepLabV3+ for segmentation, outperforms single-model alternatives on tunnel crack detection, with Score-CAM heatmaps explaining where the segmenter looks.","keywords":["Tunnel crack","Deep learning","Image classification","Semantic segmentation","Visual explanation","DenseNet-169","DeepLabV3+","Score-CAM"],"falsifier":"A concrete check has two parts. First, rerun the same training and testing protocol on NUAACrack-2000 with multiple random seeds; if DenseNet-169 does not consistently beat the six comparison classifiers, or DeepLabV3+ does not consistently beat the four segmenters, the claimed superiority is not stable. Second, apply the trained models to tunnel lining images from a different tunnel, camera, or lighting condition than NUAACrack-2000; a large drop in accuracy or pixel-overlap would show the numbers do not transfer to practice.","tokens_in":9994,"feed_emoji":"🚇","tokens_out":11904,"duration_ms":117074,"temperature":0.7,"pith_summary":"The paper sets out to make tunnel crack inspection automatic and trustworthy by splitting the problem into two steps. A DenseNet-169 classifier first separates crack images from defect-free tunnel lining images, achieving 92.23% accuracy and 39.80 frames per second on the NUAACrack-2000 dataset, ahead of six CNN and Transformer alternatives. The selected crack images then go into a DeepLabV3+ segmenter, which reaches 57.01% intersection-over-union (a pixel-overlap measure) and 67.44% F1, ahead of four segmentation baselines. The paper also uses Score-CAM heatmaps to show that the segmenter attends globally while encoding and gradually fixes on the crack itself while decoding. If these results hold, the two-stage design is a practical template for fast, explainable screening of tunnel linings.","feed_headline":"Two-stage AI pipeline spots tunnel cracks at 92% accuracy","feed_subtitle":"DenseNet-169 filters crack photos, DeepLabV3+ maps them, beating six classifiers and four segmenters.","key_machinery":"The argument runs on three named components. DenseNet-169 is a convolutional classifier whose layers receive feature maps from all preceding layers, maximizing information flow; it acts as the coarse filter that discards defect-free images. DeepLabV3+ is an encoder-decoder segmentation network built on ResNet-101, with atrous spatial pyramid pooling in the encoder using dilation rates 6, 12, and 18 to capture crack context at multiple scales, and a decoder that fuses low-level spatial features with upsampled high-level features; it is trained with dice loss to counter the imbalance between few crack pixels and many background pixels. Score-CAM is the explanation technique: it up-samples and normalizes each channel's activation map, weights the map by how much it changes the model's output relative to a baseline input, and sums the weighted maps into a heatmap. Together, the two-stage design keeps pixel-level segmentation from running on the majority of images, while Score-CAM makes the segmenter's attention inspectable.","core_discovery":"The central claim is that tunnel crack detection is best structured as a two-stage pipeline rather than as a single monolithic model. In the first stage, DenseNet-169 classifies each tunnel lining image as crack or defect-free, reaching 92.23% accuracy and 39.80 FPS; the paper reports it beats DenseNet-201, EfficientNet-B0, ResNet-50, ResNet-101, Swin Transformer, and Vision Transformer by 2.07 to 18.65 percentage points. In the second stage, DeepLabV3+ segments crack pixels in the images the classifier kept, reaching 57.01% IoU, 67.44% F1, 62.95% precision, and 82.12% recall, beating DeepLabV3, PSPNet, UNet, and UNet++. The visual-explanation experiments with Score-CAM show DeepLabV3+'s encoder starts with broad attention and its decoder narrows onto the crack, and that high-level fused features encode global context while low-level fused features encode local spatial detail.","pith_inferences":["Beyond the paper, a natural extension is to test the same two-stage pipeline on images with wet patches, shadows, or cable clutter; NUAACrack-2000 may not contain those, so transfer performance is unverified.","The Score-CAM heatmaps could be turned into a debugging tool: firms deploying crack models could inspect heatmaps on new tunnels to see whether the model is keying on crack-like stains rather than actual fractures, and retrain accordingly.","A multi-class version of the first stage could screen for leakage, spalling, and rebar exposure together, then route each flagged image to the appropriate segmenter; the paper only demonstrates binary crack screening.","The dice-loss choice suggests the authors expect crack pixels to be a small minority; a useful stress test would be measuring how both stages behave on very wide or very faint cracks that shift the pixel-class balance."],"forward_implications":["If the two-stage pipeline is adopted, crack segmentation is applied only to images already flagged as containing cracks, so computation and inspection time are concentrated where they matter.","The reported figures support the practicality of automated screening on ordinary hardware: 39.80 FPS means the classifier can process roughly forty tunnel images per second on the test setup.","DeepLabV3+'s recall of 82.12% against precision of 62.95% matches the engineering priority of not missing cracks, at the cost of extra false crack pixels.","The Score-CAM finding that the model shifts from global to crack-focused attention during decoding offers a concrete way to verify that the segmenter is responding to crack geometry rather than to unrelated image cues.","Because the dense CNN outperformed both Transformer models on this dataset, the paper implies that for small, imbalanced tunnel-image collections, CNN-based architectures remain a strong default."],"supporting_citations":[{"why":"Supplies the DenseNet architecture from which DenseNet-169, the classification stage, is taken.","marker":"Huang et al., 2018"},{"why":"Supplies the DeepLabV3+ architecture, the segmentation stage in the two-step method.","marker":"Chen et al., 2018"},{"why":"Supplies ResNet-101, the backbone used inside DeepLabV3+ for feature extraction.","marker":"He et al., 2016"},{"why":"Supplies the Score-CAM visual explanation method used to generate the heatmaps.","marker":"Wang et al., 2020"},{"why":"One of the two source references for the NUAACrack-2000 dataset, which provides the images and labels for both experiments.","marker":"Qiu et al., 2022"},{"why":"Companion NUAACrack-2000 source used to establish the crack segmentation data and ground truths.","marker":"Zhang et al., 2021"},{"why":"Supplies the threshold d=0.5 that the paper uses to decide whether a segmented crack counts as detected.","marker":"Liu and Wang, 2022"},{"why":"Supports the paper's decision to favor recall over precision in crack segmentation for engineering inspection.","marker":"Ren et al., 2020"}],"fun_headline_variants":["Two-step deep learning spots tunnel cracks at 92% accuracy","AI pipeline filters crack images, then segments them with DeepLabV3+","Two-stage pipeline beats 6 classifiers and 4 segmenters on cracks","Visual explanations reveal DeepLabV3+'s crack-detection logic"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that NUAACrack-2000, with its 512×375 tunnel lining images and ground-truth crack labels, represents the variety of lighting, crack widths, backgrounds, and camera conditions found in real tunnel inspections; if it does not, the reported accuracy and pixel-overlap figures will not transfer to practice.","fun_headline_variants_meta":{"raw":{"variants":["Two-step deep learning spots tunnel cracks at 92% accuracy","AI pipeline filters crack images, then segments them with DeepLabV3+","Two-stage pipeline beats 6 classifiers and 4 segmenters on cracks","Visual explanations reveal DeepLabV3+'s crack-detection logic"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000665,"raw_usage":{"total_tokens":3073,"prompt_tokens":1023,"completion_tokens":2050,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":639,"completion_tokens_details":{"reasoning_tokens":1982}},"tokens_in":639,"tokens_out":2050,"duration_ms":16730,"temperature":1.0,"reasoning_tokens":1982,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:11:12.368638+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete check has two parts. First, rerun the same training and testing protocol on NUAACrack-2000 with multiple random seeds; if DenseNet-169 does not consistently beat the six comparison classifiers, or DeepLabV3+ does not consistently beat the four segmenters, the claimed superiority is not stable. Second, apply the trained models to tunnel lining images from a different tunnel, camera, or lighting condition than NUAACrack-2000; a large drop in accuracy or pixel-overlap would show the numbers do not transfer to practice.","supporting_citations":[],"review_version":1}