{"id":"c82ae633-b0bd-4ef4-8c17-57cc5c052866","arxiv_id":"1908.03630","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"An adaptive, threshold-based morphological post-processing method modestly improves skin-detection masks from SegNet and a handcrafted detector on ten datasets.","lead":"This paper adds an adaptive morphological post-processing step to existing skin detectors, choosing among five operator sequences based on simple image statistics. The authors report small F-measure gains on ten public datasets, with statistical significance shown only for the SegNet base detector.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The EM threshold sentence appears to train on all test datasets; this must be resolved before the trained-morphology claims are accepted.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the reported EM threshold selection may have used the test datasets, undermining the generalization claim. The full text supports this: the methods section says EM trains on the first 2000 ECU images, but the results section says the optimal EM parameters were learned using all the datasets. This is an internal inconsistency, not merely a disagreement with consensus, and it directly affects Table 4 and the conclusion that trained morphology is better than base morphology. I do not escalate to rejection because the TM results, which use leave-one-out training, are a valid alternative path to the same qualitative claim, and the Wilcoxon test for SegNet+TM is based on those uncontaminated results. However, the ambiguity must be corrected and the EM results either justified or re-run before the paper is accepted. The paper's own Figure 3 also honestly shows cases where post-processing decreases correctly detected skin pixels, and the text admits the method cannot handle false negatives, so the main claim is appropriately modest. The proposed concrete test of re-running EM under the stated ECU-only protocol would settle whether the published EM numbers are trustworthy.","tokens_in":9728,"tokens_out":3945,"duration_ms":44982,"concrete_test":"Reproduce the EM columns of Table 4 using thresholds selected only from the first 2000 ECU images, with the ten test datasets held out, and compare the resulting F1/AP values and ranks against the published EM and BM columns. If the recomputed EM results differ materially or no longer beat BM, the sentence \"using all the datasets\" indicates test-set contamination and the EM-based conclusions must be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3 first defines EM as training on the first 2000 ECU images, then states: \"The optimal EM parameters (a1,a2,b1,b2,c1) learned for SA3 and SegNet using all the datasets are ...\". These two statements are inconsistent. If \"using all the datasets\" is literal, the EM columns in Table 4 are in-sample threshold selections on the test distribution, so the claim that EM outperforms BM and that \"a simple training on a small set of images is enough to tune our approach\" is not supported for EM. The central claim that a cheap learned post-processor generalizes therefore rests on the leave-one-out TM results alone. TM is not contaminated, and the Wilcoxon evidence for SegNet+TM over SegNet (p=0.0058) is legitimate independent support, so this is not a fatal flaw, but the ambiguity directly affects the reported EM numbers, the rank table, and the conclusion that EM and TM perform similarly. The paper must state unambiguously which datasets were used to select each threshold set, and if EM was tuned on all ten datasets, the EM results must be removed or recomputed under a genuinely held-out protocol.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a post-processing method for skin detection that applies learned morphological operators to the binary masks produced by an existing skin detector. The method defines five image classes (A through E) based on skin ratio, connected-component count, and border skin ratio, and assigns each class a fixed sequence of morphological operations. Class-assignment thresholds (a1, a2, b1, b2, c1) are tuned by grid search. The authors evaluate the post-processor on two base detectors, SegNet (deep learning) and SA3 (handcrafted), over ten public datasets from their own earlier benchmark, comparing a fixed baseline (BM), a leave-one-out trained variant (TM), and an ECU-trained variant (EM). Results are reported as per-dataset F1/AP scores, a global rank, and Wilcoxon signed-rank tests. The central claim is that the trained post-processor improves both base detectors in almost all datasets and that the trained variants outperform the fixed baseline.","tokens_in":10037,"tokens_out":2892,"duration_ms":30259,"significance":"If the main claim holds, the work offers a cheap, task-agnostic post-processing module that can raise skin-detection accuracy across diverse datasets, which would be practically valuable given the negligible computational cost (about 33 ms for a 224×224 image). The paper has concrete strengths: the evaluation uses ten heterogeneous public datasets; the code is promised open-source; and the Wilcoxon test for SegNet+TM over SegNet (p=0.0058) provides legitimate statistical evidence for the SegNet side of the claim. The leave-one-out protocol for TM is not circular, and the SegNet+TM improvement is a genuinely held-out result. However, the significance is materially weakened by (i) an explicit ambiguity about whether the EM thresholds were selected on the test distribution, and (ii) the absence of any significance test for the SA3 results, where trained morphology actually degrades performance on three of the ten datasets.","major_comments":[{"comment":"The sentence \"The optimal EM parameters (a1,a2,b1,b2,c1) learned for SA3 and SegNet using all the datasets are ...\" is directly inconsistent with the preceding description of EM as training on the first 2000 images of ECU. If \"using all the datasets\" is literal, then the EM columns in Table 4 are the result of selecting thresholds on the test distribution, making the EM-versus-BM comparison and the conclusion that \"EM and TM perform similarly\" unsupported. This is load-bearing because the paper explicitly concludes that \"a simple training on a small set of images is enough to tune our approach,\" a claim that rests on the EM results. The authors must state unambiguously which images were used to select each threshold set. If EM was tuned on all ten datasets, the EM results must be recomputed under a genuinely held-out training set (e.g., the first 2000 ECU images only) or removed.","section":"Section 3, after Table 4"},{"comment":"For the SA3 base detector, the TM variant is worse than SA3 alone or SA3+BM on three datasets (Feeval: 0.532 vs 0.539/0.540; SFA: 0.849 vs 0.863/0.849; VMD: 0.140 vs 0.147/0.140), and the paper reports no Wilcoxon test for any SA3 comparison. The claim that \"the use of morphological operators improves the performance of base approaches in almost all the datasets\" is therefore not statistically established for SA3 as a whole. At minimum, the authors should report the Wilcoxon p-values for SA3+TM vs SA3 and SA3+BM, and discuss whether the improvements on the remaining seven datasets are sufficient to compensate for the degradations, given the small number of datasets.","section":"Table 4 and statistical validation"},{"comment":"The class definitions and the morphological operator sequences in Table 3 are not fully justified. In particular, the paper states that five classes are defined but then says that classes B, C, and D receive the same operator sequence, so the actual number of distinct post-processing pipelines is three. The text should either explain what distinguishes B, C, and D in terms of the operators (e.g., the later per-component handling for A) or explicitly acknowledge that the class labels mainly drive the background-removal step for class A. As written, the reader cannot determine whether the three-way distinction among B, C, and D affects the output at all.","section":"Section 2.3 and Table 3"}],"minor_comments":[{"comment":"The abstract contains a duplicated phrase: \"is performed according to an existing skin detection approach is performed for skin segmentation.\" This should be corrected to a single clause.","section":"Abstract"},{"comment":"The caption lists \"mask obtained from SA3(τ=50)+Morphological processing\" twice; the second instance is presumably intended to be \"mask obtained from SegNet+Morphological processing.\"","section":"Figure 1 caption"},{"comment":"The sentence \"Our experiments proved that the performance of SegNet decades for low quality images\" contains a typo: \"decades\" should be \"degrades.\"","section":"Conclusions"},{"comment":"The formula for F1 is written as F1 = 2tp/(2tp + fn + fp), which omits the true negatives; this is standard, but the next sentence says \"where tn are true negatives\" even though tn does not appear in the formula. Either remove tn from the explanation or clarify that tn is not used in the F1 computation.","section":"Section 3, performance indicators"},{"comment":"The class C row lists only SR and CC features and uses \"-\" for BSR; since the text says \"Each rule is intended as the evaluation of the three conditions in AND,\" the \"-\" should be explicitly defined as \"don't care\" to avoid ambiguity.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The paper's EM threshold ambiguity is the single most important issue. It is fixable by a clarifying sentence or a recomputation, and the clean TM results for SegNet provide a solid core, so this is not a reject. However, the authors should also strengthen the statistical treatment of the SA3 results, because as it stands the paper overclaims that trained morphology improves both base detectors. I would also gently note that the benchmark framework is entirely from the authors' own earlier work; while this is not disqualifying, the paper would benefit from stating explicitly that the evaluation protocol is the one defined in [1] and that no external implementation of that protocol is used."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Loris,\n\nHere's my quick take on Lumini et al.'s morphology post-processing paper.\n\nThe useful core: they show that a cheap, rule-based adaptive morphological post-processor, switching operator sequences based on skin ratio, connected components, and border skin ratio, improves skin detection masks from two different base detectors. The adaptive rule set is a real extension of the fixed operators in Kang et al. [25]. The evaluation is broad — 10 datasets, two base detectors, three post-processing variants — and they run a proper Wilcoxon test on the main SegNet comparison, which gives p=0.0058 for SegNet+TM over SegNet. That is genuine evidence that the post-processing helps, and the leave-one-out training protocol for TM is not circular. Credit where due: this is an honest, modest step forward.\n\nNow the soft spots. First, the EM ambiguity in Section 3 is serious. They define EM as training on the first 2000 ECU images, then say \"The optimal EM parameters ... learned for SA3 and SegNet using all the datasets.\" If that second sentence is literal, the EM columns in Table 4 are selected on the test distribution, which would make the EM-vs-BM comparison and the rank table optimistic. The paper must state, unambiguously, which datasets were used to choose each threshold set. If EM was tuned on all ten, those results need to be recomputed under a held-out protocol or dropped. This is not fatal because TM is clean and supports the main claim, but it directly affects their conclusion that EM and TM perform similarly.\n\nSecond, the benefit for the handcrafted SA3 is weaker. SA3+TM worsens on Feeval, SFA, and VMD, and no significance test is reported for SA3. The overall rank improves, but the per-dataset picture is mixed. The gains are also small in absolute terms — F1 increments of 0.0 to 0.03 — so while real, they are limited in impact.\n\nThird, there are no per-dataset variances or confidence intervals, so we cannot tell how stable the improvements are. And the admitted fact that three of five classes share the same operator sequence trims the novelty further.\n\nBottom line: the paper deserves a serious referee, but the EM sentence must be fixed before the claims are accepted as written. I would ask the authors to clarify or recompute, then it is publishable in a decent venue. For me, would_cite: probably not in my own work, but I would bring it to a reading group once the ambiguity is resolved.\n\nRegards.","headline":"A modest but honest adaptive morphology post-processor; the SegNet+TM evidence is real, but the EM threshold sentence must be clarified before the trained-morphology claims hold.","tokens_in":10506,"tokens_out":2691,"would_cite":false,"duration_ms":27348,"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 learned morphological post-processing step, applied to the output of any skin detector, improves accuracy on almost all of ten diverse benchmark datasets.","keywords":["skin detection","skin segmentation","morphological operators","post-processing","convolutional neural networks","SegNet","SA3","image classification"],"falsifier":"Freeze the thresholds after training on the first 2000 ECU images only, run the identical SegNet post-processing on all ten datasets, and compare F1/AP with the unprocessed SegNet; if the post-processed version does not beat the raw detector on a clear majority of datasets, the central claim fails.","tokens_in":9546,"feed_emoji":"🖐️","tokens_out":5467,"duration_ms":52392,"temperature":0.7,"pith_summary":"This paper tries to show that skin-detection systems can be improved without retraining them, by adding a cheap post-processing stage built from trained morphological operators. The idea is to classify each raw skin mask into one of five image-pattern classes using three simple statistics, then apply a class-specific sequence of erosions, dilations, fillings, and small-object removals that strip away false-positive background pixels. The authors test the post-processor on top of a deep-learning detector (a fine-tuned SegNet) and a handcrafted detector (SA3), across ten public datasets. Their central claim is that trained morphology outperforms a fixed morphological baseline and no post-processing, with the best combination ranking first among all compared approaches.","feed_headline":"Learned morphology lifts skin detection on nearly all 10 datasets","feed_subtitle":"A five-class rule picks per-image morphological cleanup; SegNet plus trained operators takes first place in the benchmark.","key_machinery":"The central mechanism is a five-class decision rule plus class-specific morphological operator sequences. For a binary mask $BW$, the method computes the skin ratio $SR$ on a hole-filled copy; if $SR$ exceeds threshold $a_1$, the mask is heavily eroded and the number of connected components $CC$ is counted, and if $CC$ is below $b_1$, the border skin ratio $BSR$ decides between class A and class B. Additional thresholds $a_2$ and $b_2$ separate class D from the default class E. Each class then receives a fixed chain of morphological commands, with the structural element a disk of radius 6, and the final mask is multiplied pixel-wise with the original segmentation. This mechanism is what carries the claimed gain: it converts a one-size-fits-all post-processing into a per-image adaptive cleanup at negligible computational cost.","core_discovery":"On the paper's own terms, the discovery is that false-positive errors in skin segmentation are structured enough to be corrected by a small set of learned image-processing rules. Each binary output mask is described by its skin ratio, connected-component count, and border skin ratio; these three features place the mask into classes A through E, and a lookup table then applies a fixed operator sequence per class, with class A (background classified as skin) receiving an extra step that removes the largest connected component. The thresholds for the rules are obtained by grid search, either on the ECU training set or by leave-one-out over the other datasets. In the reported experiments the trained versions (TM and EM) beat the base morphological version (BM), which in turn beats the raw detectors on almost all datasets; the Wilcoxon signed-rank test gives SegNet+TM over SegNet a p-value of 0.0058.","pith_inferences":["Because the post-processor reads only binary masks and three scalar statistics, a natural untested extension is to apply the same scheme to other binary segmentation outputs where false positives from background are the dominant error.","The paper itself notes that working from probability masks instead of binary masks could recover false negatives; a testable follow-up would feed soft masks through the same class rules with several thresholds.","The reported sentence that optimal EM parameters were learned 'using all the datasets' leaves open whether the comparison is fully held out; a strict holdout replication is the cleanest way to confirm the transfer claim."],"forward_implications":["Any current skin detector can be upgraded by appending this post-processor, with no retraining of the underlying model and a runtime of roughly 0.03 seconds per 224x224 image.","The class-based formulation means the thresholds can be re-learned on a small labeled set, so the method should adapt to a new domain cheaply.","Gains concentrate on false-positive reduction; detectors that mostly miss skin regions will not be helped, because the post-processor cannot recover false negatives.","Statistical support from the Wilcoxon test suggests the improvement over both the raw detector and the fixed morphological baseline is not a single-dataset accident."],"supporting_citations":[{"why":"Defines the ten-dataset testing framework and the baselines used for comparison.","marker":"[1]"},{"why":"Supplies the SegNet encoder-decoder architecture that is fine-tuned as the deep-learning detector.","marker":"[16]"},{"why":"Supplies the SA3 adaptive handcrafted skin-segmentation method used as the second baseline.","marker":"[21]"},{"why":"Provides the ECU dataset whose first 2000 images train both SegNet and the EM thresholds.","marker":"[19]"},{"why":"Provides the fixed morphological operator baseline (BM) adapted in this work.","marker":"[25]"},{"why":"Supplies the Wilcoxon signed-rank test used to claim statistical significance.","marker":"[36]"}],"fun_headline_variants":["Learned morphology rules beat raw skin detectors","Trained morphology cleanup lifts skin detection","Learned morphology postprocessing refines skin masks","Rule-based morphology beats raw skin segmentation","Trained morphology rules improve skin detection on most datasets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole improvement rests on the five grid-search thresholds $(a_1,a_2,b_1,b_2,c_1)$ transferring from the training set to every test dataset; if those thresholds were tuned with any access to test data, or if they do not transfer to a new domain, the reported gains would shrink or vanish.","fun_headline_variants_meta":{"raw":{"variants":["Learned morphology rules beat raw skin detectors","Trained morphology cleanup lifts skin detection","Learned morphology postprocessing refines skin masks","Rule-based morphology beats raw skin segmentation","Trained morphology rules improve skin detection on most datasets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001581,"raw_usage":{"total_tokens":6226,"prompt_tokens":786,"completion_tokens":5440,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":402,"completion_tokens_details":{"reasoning_tokens":5373}},"tokens_in":402,"tokens_out":5440,"duration_ms":36300,"temperature":1.0,"reasoning_tokens":5373,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:06:50.590013+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Freeze the thresholds after training on the first 2000 ECU images only, run the identical SegNet post-processing on all ten datasets, and compare F1/AP with the unprocessed SegNet; if the post-processed version does not beat the raw detector on a clear majority of datasets, the central claim fails.","supporting_citations":[{"cited_title":"Lumini, L","cited_arxiv_id":null,"evidence_quote":"Defines the ten-dataset testing framework and the baselines used for comparison."},{"cited_title":"Kawulok, J","cited_arxiv_id":null,"evidence_quote":"Supplies the SA3 adaptive handcrafted skin-segmentation method used as the second baseline."},{"cited_title":"Phung, A","cited_arxiv_id":null,"evidence_quote":"Provides the ECU dataset whose first 2000 images train both SegNet and the EM thresholds."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the fixed morphological operator baseline (BM) adapted in this work."},{"cited_title":"Demšar, Statistical Comparisons of Classifiers over Multiple Data Sets, J","cited_arxiv_id":null,"evidence_quote":"Supplies the Wilcoxon signed-rank test used to claim statistical significance."}],"review_version":1}