{"id":"ae7e1a7c-cf94-464f-a202-f1241330e678","arxiv_id":"2608.12725","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Using segmentation-guided inpainting to remove co-occurring defects from electroluminescence images improves multi-label photovoltaic defect classification, with the largest gains on rare classes and low-data regimes.","lead":"This paper introduces Generative Defect Isolation, a method that uses image inpainting to erase selected defects from solar cell photos and create clean single-defect training examples. It reports consistent improvements in multi-label defect classification, especially for rare defect types and limited training data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline gains of +63.6% and 26% are from test-set-threshold-tuned runs; no fixed-threshold class-level or co-occurrence analysis is reported, so the mechanism claim may rest on threshold overfitting to tiny samples.","rationale":"The reader correctly flagged test-set threshold tuning and tiny rare-class samples as risks, and I agree these are the load-bearing weak point. My concern is more specific than the reader's primary weakest_assumption (LaMa inpainting realism): the fixed-threshold multi-seed study already gives reasonable evidence that GDI helps on aggregate metrics at the 20% split, so the inpainting-realisticity concern is partially addressed. What remains untested is whether the headline class-level and co-occurrence numbers survive without test-set threshold selection. That is a concrete, checkable gap. Because the paper already discloses the threshold protocol and provides a fixed-threshold robustness study, the appropriate verdict remains CONDITIONAL: accept the aggregate claim conditional on the class-level and co-occurrence claims being re-evaluated at a fixed or validation-tuned threshold. I do not see grounds to move to ACCEPT or REJECT from the current conditional status.","tokens_in":18994,"tokens_out":5262,"duration_ms":55137,"concrete_test":"Rerun the 20% split for EfficientNetV2-L, ViT-S, and ViT-L with the same four seeds (24, 42, 67, 76), evaluating at a fixed threshold tau=0.5 (and, as a second check, at thresholds chosen on a held-out validation split rather than the test split). Report the class-wise F1 for Contact_BeltMarks, the total off-diagonal co-occurrence error count from Figure 6, and their per-seed distributions. If the +63.6% class gain and the 26% error reduction are not consistently reproduced at the fixed threshold, the headline mechanism claims should be reported as threshold-dependent rather than as intrinsic properties of GDI.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanism claim—that GDI resolves co-occurrence ambiguity—is carried by two headline numbers from the 20% split: the +63.6% F1 gain for Contact_BeltMarks and the 26% reduction in co-occurring errors (Section 4.6, Figures 5 and 6). Both are produced under the Section 4.2.2 protocol, where per-class thresholds are chosen to maximize macro F1 on the test set. Contact_BeltMarks has only 5 test samples and 3 inpainted training samples in this split (Table 3), so threshold selection can fit noise; a relative F1 difference can be driven by a handful of label flips. The fixed-threshold, multi-seed experiment in Section 4.9 provides genuine support for aggregate gains at 20%, but it reports only macro F1 and zero-one accuracy. It does not report class-wise F1 or error co-occurrence matrices. Thus the specific evidence for the rare-class and co-occurrence-disambiguation parts of the central claim has not been separated from test-set threshold overfitting. If, at a fixed threshold, the +63.6% gain and the 26% reduction vanish or are inconsistent across seeds, the mechanism-level claim is unsupported even though the augmentation may still help on aggregate metrics.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper introduces Generative Defect Isolation (GDI), an annotation-guided data augmentation technique for multi-label defect classification in electroluminescence (EL) images of photovoltaic cells. GDI uses the LaMa inpainting model to remove selected defects from multi-defect training images, generating single-defect and no-defect training samples from existing segmentation annotations. The authors evaluate GDI on the UCF-EL-Defect dataset with ViT-S, ViT-L, and EfficientNetV2-L classifiers, reporting that GDI improves zero-one accuracy and macro F1 across several training-data fractions, with the largest relative gains in low-data regimes. They further report a 63.6% F1 gain for the rare Contact_BeltMarks class and a 26% reduction in co-occurring error pairs, and they compare against copy-paste augmentation and re-implemented baselines from prior work. The main experiments select per-class decision thresholds on the test set, a fact disclosed in Section 4.2.2, and Section 4.9 provides a fixed-threshold, multi-seed robustness check for aggregate metrics.","tokens_in":19180,"tokens_out":16005,"duration_ms":139049,"significance":"GDI is a practical and clearly described idea: repurposing pixel-level segmentation annotations to create cleaner training examples for a downstream classification task, with no additional inference cost. The paper is honest about the test-set threshold selection in the main protocol and includes a complementary fixed-threshold multi-seed experiment that supports the aggregate claim that GDI improves mean macro F1 and zero-one accuracy across all three architectures. The comparison with copy-paste augmentation and the inclusion of code and synthetic images are useful contributions. However, the stress-test concern is valid: the headline rare-class and co-occurrence-disambiguation numbers come from threshold-tuned single runs, and the fixed-threshold study does not cover class-wise or co-occurrence metrics. The mechanism-level claims therefore need additional support before the paper can be accepted as stated.","major_comments":[{"comment":"The +63.6% relative F1 gain for Contact_BeltMarks is computed under the main protocol in which per-class thresholds are selected to maximize macro F1 on the test set, and this class has only 5 test samples and 3 inpainted training samples in the 20% split. Threshold selection can therefore move F1 by a large relative amount based on one or two prediction changes, and the fixed-threshold experiment in Section 4.9 reports only aggregate metrics, not class-wise F1. Please report class-wise F1 at the fixed threshold tau=0.5 across the four seeds, or with thresholds chosen on a validation split, so that the rare-class claim is separated from threshold overfitting.","section":"Section 4.6, Table 3, Figure 5"},{"comment":"The 26% reduction in co-occurring error pairs and the specific pair-level reductions (e.g., Unknown/Contact_NearSolderPad falling from 165 to 66) are also obtained from the test-set-threshold-tuned run. Since the co-occurrence analysis is the main evidence for the claimed mechanism of resolving learning ambiguity, it should be recomputed at the fixed threshold (tau=0.5) or with validation-based thresholds for all seeds. Without that, the aggregate-only robustness check in Section 4.9 cannot support the co-occurrence-disambiguation part of the central claim.","section":"Section 4.6, Figure 6"},{"comment":"The definition of an 'error' in the co-occurrence matrix is ambiguous. The text states that an off-diagonal cell counts samples where the model made an error on both classes, but the subsequent example of the model 'co-predicting' Unknown on Contact_NearSolderPad images suggests false positives alone. Please specify whether the matrix counts false positives, false negatives, or both, and make the example consistent with that definition, because the 26% figure and the 60% reduction for the Unknown/Contact_NearSolderPad pair are headline quantitative claims.","section":"Section 4.6, Figure 6"}],"minor_comments":[{"comment":"The sum of the upper triangle of the baseline matrix as printed is 1759, not 1774 as stated in the text; please verify the printed matrix or the reported sum.","section":"Section 4.6, Figure 6"},{"comment":"The abstract and conclusion repeat the +63.6% and 26% figures without noting that they come from the test-set-threshold-tuned protocol; consider adding that caveat wherever these numbers appear.","section":"Section 4.2.2 / Abstract / Section 6"},{"comment":"The comparison with prior work re-implements their backbones under the authors' 12-class label space and a new 80/20 split; the 'new state-of-the-art' claim should be framed as relative to these re-implemented baselines rather than to the original published results.","section":"Section 4.7, Table 4"},{"comment":"The sentence 'The network is trained using a combined objective function...' describes the pre-training of LaMa, not training performed in this paper; rephrase to avoid implying that the authors trained the inpainting model.","section":"Section 3.2.3"},{"comment":"The phrase '100% drop in accuracy' is unclear; state the actual accuracy values and describe the change as a relative change of -100%.","section":"Section 4.3"},{"comment":"The 10% area filter and the 15x15 dilation kernel are chosen by qualitative inspection; a sensitivity analysis would make the method's dependence on these hyperparameters explicit.","section":"Section 3.2.1 / Section 3.2.2"}],"recommendation":"major_revision","confidential_remarks":"To the editor: This is a solid empirical paper with a well-disclosed main protocol, but the two headline quantitative claims (rare-class F1 gain and co-occurrence-error reduction) rest on test-set-threshold-tuned single runs. The authors' own Section 4.9 shows they know how to run a fixed-threshold protocol; requiring them to extend that protocol to the class-wise and co-occurrence metrics is a tractable revision. The paper is within scope for Solar Energy and the data-centric idea is useful. I recommend major revision rather than rejection because the central augmentation claim is supported for aggregate metrics."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: GDI (inpaint away the other defects to make clean single-defect training samples) is a real idea, and the paper has honest evidence that it helps. The fixed-threshold, multi-seed experiment at 20% shows GDI improving mean macro F1 and zero-one accuracy for all three architectures, and it wins in 11 of 12 paired accuracy runs and 10 of 12 paired F1 runs. That is the strongest part of the paper, and it is genuinely good.\n\nWhat's new: the subtractive, mask-guided inpainting to repurpose segmentation data is distinct from GAN synthesis or copy-paste. Composing it with LaMa is straightforward, and the comparison against copy-paste, with physical reasoning about why copy-paste can hurt, is a useful contribution. The qualitative inpainting examples are convincing, and the failure mode discussion (Figure 8) is honest.\n\nWhere it's softer: the two headline numbers—+63.6% rare-class F1 and 26% co-occurrence error reduction—come from the main protocol where per-class thresholds are chosen on the test set. Contact_BeltMarks has 5 test samples and 3 inpainted training samples, so that F1 delta is close to noise. The fixed-threshold experiment does not report class-wise F1 or co-occurrence matrices, so the mechanism-level claims are not yet separated from threshold overfitting. The authors disclose the setup and include the fixed-threshold check, so I am not accusing them of hiding anything; they just headline numbers that are the least robust ones in the paper.\n\nMinor: the ViT-S 1% zero-accuracy run is dismissed with a footnote about the baseline being near zero, but a total collapse still deserves a sentence of explanation. The code link has no commit hash, and the training hyperparameters are not fully specified. The \"new benchmark\" claim is based on their own reimplementation of the comparators, which is reasonable but should be stated more carefully.\n\nBottom line: the central aggregate claim holds up; the mechanism-level claims need a fixed-threshold class-wise analysis. This paper deserves a serious referee—send it out, ask for the extra analysis and error bars on rare classes. I would cite it if I worked on data-centric augmentation or segmentation-to-classification repurposing.","headline":"GDI is a genuinely useful data-centric augmentation with aggregate gains that survive a fixed-threshold check, but the headline rare-class and co-occurrence numbers come from test-set threshold tuning and are less solid than the abstract suggests.","tokens_in":19766,"tokens_out":2753,"would_cite":true,"duration_ms":28360,"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":"This paper shows that Generative Defect Isolation (GDI) — erasing non-target defects via LaMa inpainting — improves multi-label photovoltaic defect classification, raising rare-class F1 by up to 63.6% and cutting co-occurring errors by 26%.","keywords":["Photovoltaics","Electroluminescence imaging","Multi-label defect classification","Synthetic data","Inpainting","Generative Defect Isolation","Data augmentation"],"falsifier":"Filter out every GDI-generated no-defect image whose inpainted region fails to reconstruct grid lines (the Figure 8-type case) and retrain the same classifiers; if the macro F1 and zero-one accuracy gains over baseline persist, the benefit comes from defect isolation itself, whereas if they collapse to noise, the claimed gains were carried by inpainting fidelity rather than by removing co-occurrence ambiguity.","tokens_in":18723,"feed_emoji":"☀️","tokens_out":10963,"duration_ms":104591,"temperature":0.7,"pith_summary":"The paper asks whether the ambiguity of multi-label defect classification in electroluminescence (EL) images of solar cells can be reduced by changing the training data rather than the model. It proposes Generative Defect Isolation (GDI): for images with several co-occurring defects, the ground-truth segmentation masks are used to erase all but one defect with the LaMa inpainting model, producing clean single-defect examples and one defect-free example from each source image. Across ViT-S, ViT-L, and EfficientNetV2-L, training on the original plus GDI-augmented images improves zero-one accuracy and macro F1 in nearly every data-scarce setting, with the largest gains on rare classes (up to +63.6% F1) and a 26% reduction in co-occurring classification errors. The authors position GDI as a way to repurpose expensive segmentation annotations into faster, production-ready multi-label classifiers at no inference-time cost.","feed_headline":"Inpainting away defects lifts rare solar-defect F1 by 63.6%","feed_subtitle":"It turns multi-defect solar images into clean single-defect training examples, cutting co-occurrence errors by 26%.","key_machinery":"The central object is the GDI pipeline built on LaMa (Large Mask Inpainting), a neural inpainting network whose bottleneck uses Fast Fourier Convolution residual blocks: one branch processes local texture with ordinary convolutions while the other applies a real FFT, a $1\\times 1$ convolution, and an inverse FFT to obtain an image-wide receptive field that can reproduce the periodic grid lines of PV cells. GDI selects a target defect only if its annotated area exceeds 10% of the image, dilates the mask of all other defects with a $15\\times 15$ kernel, and asks LaMa to fill the masked region, then repeats for each qualifying defect and finally generates a no-defect sample by inpainting all defects at once. The augmented single-defect images are one-hot labeled and added to the original multi-defect training set, so the classifier sees both clean single-defect mappings and real co-occurrence patterns.","core_discovery":"GDI is the claim that annotating which pixels belong to each defect lets a generative inpainting model subtract defects from real EL images, yielding training examples in which one defect appears alone or no defect appears at all, and that training on this hybrid set disentangles visual features better than training on the original multi-defect images alone. In the paper's strongest results, GDI raises macro F1 from 0.7672 to 0.7744 and zero-one accuracy from 0.5943 to 0.6046 for EfficientNetV2-L on the full dataset, and at the 20% data split it lifts rare-class F1 by 63.6% (Contact_BeltMarks) and 20.1% (Crack_Isolated). The error-co-occurrence matrices show the total number of paired class errors falling from 1,774 to 1,312, a 26% drop, with the Unknown-class confusions decreasing most sharply; multi-seed experiments at a fixed threshold of $\\tau=0.5$ confirm the gains are not an artifact of per-class threshold tuning.","pith_inferences":["Going beyond the paper's experiments, GDI should transfer to other defect domains where segmentation masks already exist and the background has strong structural regularity — pipe welds, textiles, or metal surfaces — because the core mechanism, removing confounded co-occurring labels from training images, is not specific to photovoltaics.","The fixed-threshold multi-seed results (11 of 12 accuracy improvements, 10 of 12 F1 improvements) suggest a practical deployment rule the paper does not state: apply GDI before searching over architectures, since the smallest models receive the largest relative lift.","A testable extension is to filter out the acknowledged grid-line failure cases (Figure 8) from the generated No_Defect set and retrain; if the gains persist, defect isolation itself is the driver, whereas if they shrink, inpainting fidelity is the true source of the improvement.","The authors name a domain-specific inpainting model pretrained on PV cell topology as future work; if such a model eliminates the grid-line failure mode, the likely downstream effect is a further drop in Unknown-class confusion rather than a large change in already-well-classified common defects."],"forward_implications":["Rare defect classes become learnable from just a handful of clean, inpainted single-defect examples: Contact_BeltMarks, with only 6 original labels in the 20% split, gains +63.6% relative F1 after 3 GDI samples are added.","GDI's benefit is largest when training data is scarce — e.g., +125.1% zero-one accuracy for ViT-S at 10% data — and shrinks but stays positive at full data, making the method most valuable where annotation budgets are tight.","Because the Unknown class gains +16.5% F1 without any new Unknown samples, GDI indirectly reduces a model's overuse of ambiguous labels by drawing clearer boundaries for the known defect classes.","GDI is a one-time offline preprocessing step with no inference-time cost, so the expensive pixel-level segmentation knowledge is distilled offline into a fast classifier that can screen and bin cells on a production line.","GDI outperforms Copy-Paste augmentation on all three architectures, and the paper attributes this to the physical implausibility of pasting defects whose locations are constrained by busbar stress and contact interfaces."],"supporting_citations":[{"why":"Supplies the segmentation-annotated EL dataset whose pixel-level masks drive GDI's mask generation and whose test set is used for all evaluations.","marker":"[8]"},{"why":"LaMa is the inpainting architecture with Fast Fourier Convolutions that GDI uses to erase non-target defects and restore periodic PV structure.","marker":"[11]"},{"why":"Provides the implementation of the inpainting model used to generate the GDI augmented samples.","marker":"[22]"},{"why":"Copy-Paste augmentation is the additive comparison method that GDI must beat to show its gains come from subtraction rather than added diversity.","marker":"[33]"},{"why":"Supplies prior MobileNetV2 and Xception baselines that are re-evaluated with and without GDI for model-agnostic comparison.","marker":"[30]"},{"why":"EfficientNetV2 is one of the three architectures tested; its Large variant achieves the paper's top full-data F1 and accuracy.","marker":"[27]"},{"why":"The Vision Transformer architecture family for ViT-S and ViT-L, whose multi-label results under GDI are reported.","marker":"[24]"},{"why":"Defines the ViT-S configuration used in the study, distinguishing it from the larger ViT-L.","marker":"[26]"}],"fun_headline_variants":["Inpainting solar defects boosts rare-class F1 by up to 63.6%","Single-defect training from multi-defect images cuts errors 26%","Generative defect isolation: rare F1 +63.6%, co-occurrence errors -26%","Rare solar defects: F1 up 63.6% by inpainting them away"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"GDI's gains depend on LaMa, pretrained on natural images, producing inpainted EL-cell regions that are visually and structurally faithful enough that the clean single-defect and no-defect examples transfer to real test cells — the paper itself acknowledges a failure mode when combined defect masks hide all grid lines.","fun_headline_variants_meta":{"raw":{"variants":["Inpainting solar defects boosts rare-class F1 by up to 63.6%","Single-defect training from multi-defect images cuts errors 26%","Generative defect isolation: rare F1 +63.6%, co-occurrence errors -26%","Rare solar defects: F1 up 63.6% by inpainting them away"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001521,"raw_usage":{"total_tokens":6111,"prompt_tokens":982,"completion_tokens":5129,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":598,"completion_tokens_details":{"reasoning_tokens":5035}},"tokens_in":598,"tokens_out":5129,"duration_ms":35771,"temperature":1.0,"reasoning_tokens":5035,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:38:45.711628+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Filter out every GDI-generated no-defect image whose inpainted region fails to reconstruct grid lines (the Figure 8-type case) and retrain the same classifiers; if the macro F1 and zero-one accuracy gains over baseline persist, the benefit comes from defect isolation itself, whereas if they collapse to noise, the claimed gains were carried by inpainting fidelity rather than by removing co-occurrence ambiguity.","supporting_citations":[{"cited_title":"Fioresi, D","cited_arxiv_id":null,"evidence_quote":"Supplies the segmentation-annotated EL dataset whose pixel-level masks drive GDI's mask generation and whose test set is used for all evaluations."},{"cited_title":"Suvorov, E","cited_arxiv_id":null,"evidence_quote":"LaMa is the inpainting architecture with Fast Fourier Convolutions that GDI uses to erase non-target defects and restore periodic PV structure."},{"cited_title":"Abdelsattar, A","cited_arxiv_id":null,"evidence_quote":"Supplies prior MobileNetV2 and Xception baselines that are re-evaluated with and without GDI for model-agnostic comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"EfficientNetV2 is one of the three architectures tested; its Large variant achieves the paper's top full-data F1 and accuracy."}],"review_version":1}