acceptodds
Under review as a conference paper at ICLR 2027

WAVELET-GUIDED BRANCH-CONDITIONED GROUPED NORMALIZING FLOW FOR POWER-LINE ANOMALY DETECTION

Abstract

UAV power-line inspection requires anomaly detectors to distinguish equipment faults from substantial changes in background, viewpoint, and illumination. We propose Wavelet-Guided Branch-Conditioned Grouped Normalizing Flow(WBGFlow), a shared conditional density model that organizes complementary observations for this setting. RAW appearance, low-frequency structure, and high frequency energy retain distinct feature groups throughout the flow. Per-block identity embeddings adapt shared coupling subnetworks to each observation source. All three branches learn normality jointly, while the final score selects RAW and LL latent energies. This connects complementary-observation training with explicit control over anomaly-ranking evidence. On InsPLAD-fault, WBGFlow achieves the highest mean image-level AUROC among the compared methods at 94.62%, averaged over two training seeds, exceeding DifferNet by 2.16 percentage points in macro-average with gains in all five evaluated categories. A complementary MVTecAD evaluation yields 95.78%. Component ablations examine the roles of grouping and identity conditioning, while fixed-model scoring analysis supports separating training observations from ranking contributions. Together, the results demonstrate the value of identity-preserving shared density modeling for uncontrolled power-line inspection.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Unsupervised Anomaly Detection, Normalizing Flows, Wavelet Decomposition, Conditional Density Estimation, Uncontrolled Environments, Power-Line Inspection

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_rEz6OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes WBGFlow, an anomaly detection method for UAV power-line inspection that combines wavelet-based frequency decomposition with grouped, branch-conditioned normalizing flows. The method achieves competitive results on InsPLAD-fault and improves upon DifferNet. Additional ablation studies investigate the individual contributions of grouping, branch conditioning, and score selection to anomaly detection performance.

Critical strengths
  • The paper addresses a relevant practical problem: anomaly detection under varying acquisition conditions.
  • The proposed architecture combines existing components in a potentially useful way.
  • The method improves upon DifferNet on InsPLAD-fault, and the authors provide component ablations.
Critical weaknesses

Weaknesses

  1. Limited methodological novelty. The method primarily combines established techniques, including frequency decomposition and conditional normalizing flows. While such a combination can constitute a valid contribution, the paper does not introduce a substantially new methodological principle. Its contribution therefore rests primarily on the empirical benefits of the proposed combination.

  2. Insufficient validation of the central motivation. The paper emphasizes robustness to changes in background, viewpoint, and illumination. However, the experiments do not directly evaluate robustness to these factors. In particular, comparisons against simpler approaches using appropriate data augmentation would help establish whether the proposed architectural complexity is necessary.

  3. Limited improvement over stronger existing methods. Although WBGFlow improves upon DifferNet on InsPLAD-fault (94.62% versus 92.46% image-level AUROC), its advantage becomes marginal when compared with stronger existing methods. For example, AttentDifferNet [3] reports 94.34%, while a more recent frequency-based approach [4] reports 94.70%. Furthermore, CLAdapter [5] reports substantially higher performance (98.29%), although it relies on pretrained foundation models and therefore represents a different methodological setting. Similarly, WBGFlow's 95.78% AUROC on MVTec AD is substantially below the approximately 99% achieved by stronger contemporary methods [1], [2]. These comparisons weaken the evidence that the proposed framework provides a meaningful advance over existing approaches.

References

[1] Q. Chen, H. Luo, C. Lv, and Z. Zhang, “A unified anomaly synthesis strategy with gradient ascent for industrial anomaly detection and localization,” in Proc. Eur. Conf. Comput. Vis., vol. 15125, 2024, pp. 37–54.

[2] A. Bauer, S. Nakajima, and K.-R. Müller, “Self-supervised autoencoders for visual anomaly detection,” Mathematics, vol. 12, no. 24, p. 3988, 2024.

[3] A. L. Vieira e Silva, F. Simões, D. Kowerko, T. Schlosser, F. Battisti, and V. Teichrieb, “Attention modules improve image-level anomaly detection for industrial inspection: A DifferNet case study,” in Proc. IEEE/CVF Winter Conf. Appl. Comput. Vis. (WACV), 2024, pp. 8246–8255.

[4] J. Wei, X. Liu, H. Ren, and X. Xu, “Unsupervised anomaly detection via multi-frequency structure and visual sampling,” in Proc. 2nd Int. Conf. Image Process., Multimedia Technol. Mach. Learn. (IPMML), 2025, pp. 245–249.

[5] Q. Li, F. He, H. Chen, X. Ning, K. Wang, and Z. Wang, “Unleashing foundation vision models: Adaptive transfer for diverse data-limited scientific domains,” in Adv. Neural Inf. Process. Syst. (NeurIPS), vol. 38, 2025.

Questions for authors

Questions for the Authors

  1. Rationale behind the architectural choices: The paper describes several architectural choices, including frequency separation, grouped coupling, and shared conditional normalizing flows, but their underlying rationale remains unclear. For example, separating frequency components may improve the interpretability of individual branches, but why should it improve anomaly detection when neural networks can learn to distinguish such features internally? It would be helpful if the authors could clarify why these design choices are expected to be beneficial beyond the empirical results of the ablation studies.

  2. Robustness to acquisition conditions: Since robustness to variations in background, viewpoint, and illumination is a central motivation, could the authors provide evidence that WBGFlow is less sensitive to these variations than existing methods?

  3. Comparison with state-of-the-art methods: The evaluation would benefit from comparisons with stronger contemporary anomaly detection methods on both InsPLAD-fault and MVTec AD. In particular, how does WBGFlow compare with AttentDifferNet [3], the frequency-based approach [4], and CLAdapter [5] on InsPLAD-fault, as well as GLASS [1] and reconstruction-based approaches [2] on MVTec AD? Please discuss differences in training assumptions where relevant.

7 comments
AuthorsOpenReview ↗

Response to Reviewer 3 W1

Thank you for reviewing our paper and recognizing the importance of anomaly detection in uncontrolled power-line inspection and WBGFlow's improvement over DifferNet. We respond to each point below.

Weakness 1

Thank you for the feedback. We agree that frequency decomposition and conditional normalizing flows are established techniques. This work introduces a mechanism for shared, branch-conditioned density modeling across different frequency observations.

WBGFlow uses Grouped Coupling to preserve distinct feature groups for RAW, LL, and HF, so each branch retains its own latent representation through the invertible transformations. A Branch Identity Embedding in each coupling block identifies the observation to the shared subnetworks, allowing the same parameters to produce branch-specific affine transforms. WBGFlow shares coupling parameters across the three branches for joint density modeling while keeping their feature groups separate. This lets the branches share one flow instead of requiring a separate flow for each frequency.

The existing ablations show what each of these mechanisms contributes. On Upper Shackle in Table 3 of the original manuscript, Grouped Flow raises AUROC from 93.81% to 94.41%. Adding Branch Identity Embedding increases AUROC to 95.22%, and selective RAW+LL scoring raises it to 96.14%.

Using the same counting convention, WBGFlow has 150.427M total parameters, 13.82% fewer than DifferNet's 174.559M. In the expanded 5-seed evaluation on InsPLAD-fault, WBGFlow's macro-average AUROC exceeds DifferNet's by 2.27 percentage points.

WBGFlow improves on DifferNet with fewer parameters. The improvement is associated with how the frequency observations are organized and how their densities are modeled conditionally.

AuthorsOpenReview ↗

Response to Reviewer 3 W2

Weakness 2

Thank you for the suggestion. To directly examine how acquisition changes affect detection, we added controlled imaging corruptions and a background evaluation on InsPLAD-fault. We also report the results for random in-plane rotation from Appendix Table B.2.

To address this point, we added a controlled acquisition corruption experiment.

Controlled imaging corruptions and matched training augmentation. For each training configuration, we used the same checkpoint, test split, and scoring rule to evaluate the original and corrupted images. We applied brightness changes, contrast reduction, Gaussian blur, and Gaussian noise to the same test images. To measure how much matched training augmentation can narrow the performance gap, we also evaluated DifferNet trained with that augmentation.

Table 3.1. Comparison with DifferNet trained with matched augmentation

Test conditionDifferNetDifferNet + Aug.WBGFlowΔWBG−Aug\Delta_{\mathrm{WBG}-\mathrm{Aug}}DifferNet dropWBGFlow drop
Clean89.6090.1993.97+3.780.000.00
Brightness change(mean dark/light, severity2)86.2988.5392.28+3.753.311.69
Contrast reduction (0.70)83.7787.4692.54+5.085.831.43
Gaussian blur (radius1.0)85.4585.9192.91+7.004.151.06
Gaussian noise (std0.03)58.5058.5664.15+5.5931.1029.82
Mean corrupted AUROC78.5080.1285.47+5.3611.108.50
Mean AUROC drop (pp)11.1010.088.50-1.58––

Each condition reports macro-average image-level AUROC (%) across the five equipment categories. DifferNet + Aug. denotes DifferNet trained with matched augmentation. The brightness result averages darkening and brightening at the same severity level. Drop is the difference between a method's AUROC on the original and corrupted images, in percentage points (pp). The mean corrupted AUROC and mean drop weight the four corruption types equally. The difference column subtracts the augmented DifferNet result from the WBGFlow result. Summary values and differences are calculated from the per-condition values shown and then rounded to two decimal places.

Matched training augmentation raises DifferNet's mean AUROC across the four corruptions from 78.50% to 80.12%. WBGFlow achieves higher AUROC than augmented DifferNet on all four corruptions, with a mean of 85.47%. Relative to each model's AUROC on the original images, the mean drop is 8.50 percentage points for WBGFlow and 10.08 for augmented DifferNet. With the augmentation used in this experiment, training augmentation narrows the gap, and WBGFlow retains its detection advantage.

Background changes. We randomly sampled 300 normal and anomalous images from the InsPLAD-fault test set and used manual background annotations to construct the background comparison.

Table 3.2. Detection performance on the InsPLAD-fault background evaluation subset

MethodClean AUROC (%)Dark-background AUROC (%)AUROC drop (pp) ↓
DifferNet87.7177.6010.11
WBGFlow94.1684.989.18

The results use 300 randomly sampled images from the InsPLAD-fault test set, including normal and anomalous samples. Both methods were evaluated on the same images and under the same background conditions. The AUROC drop is the AUROC under the original condition minus the AUROC under the dark-background condition, in percentage points (pp).

As shown in Table 3.2, WBGFlow achieves 84.98% image-level AUROC under dark backgrounds, 7.38 percentage points higher than DifferNet's 77.60%. Relative to their results on the original images, WBGFlow and DifferNet drop by 9.18 and 10.11 percentage points, respectively. WBGFlow maintains a higher detection AUROC with a smaller drop under the tested dark-background condition.

In-plane rotation. Appendix Table B.2 evaluates both methods with fixed checkpoints and matched random rotation sequences. With a single randomly rotated view (T=1), WBGFlow and DifferNet achieve five-category macro-average AUROC of 91.488% and 83.354%, respectively, a difference of 8.134 percentage points. WBGFlow achieves higher macro-average AUROC at all six tested view budgets. This result quantifies detection performance under random in-plane rotation.

We will include these results in the revision and limit our robustness conclusions to the conditions covered by this evaluation.

AuthorsOpenReview ↗

Response to Reviewer 3 W3

Weakness 3

Thank you for raising this point. We extended WBGFlow's evaluation from two to five independent training seeds. The five runs yield a mean macro-average image-level AUROC of 94.73%, with a sample standard deviation of 0.25 percentage points across seeds.

1. Comparison with AttentDifferNet

AttentDifferNet-SE reports an average AUROC of 94.34% on InsPLAD-fault. WBGFlow's five-seed mean is 94.73%, 0.39 percentage points higher. We also compared model parameter counts. In our implementations based on the paper's architecture descriptions, AttentDifferNet-SE and AttentDifferNet-CBAM contain 174.572M and 174.573M parameters, respectively, compared with 150.427M for WBGFlow. The counts for the two AttentDifferNet variants come from our unofficial implementations of the published architectures. WBGFlow achieves competitive detection performance with a more compact model.

2. Comparison with Multi-Frequency Structure and Visual Sampling

Table I of the Multi-Frequency paper reports a macro-average AUROC of 94.7%. WBGFlow's five-seed mean is 94.73%, which also rounds to 94.7% at the same one-decimal precision. The category results in Table I show that WBGFlow achieves higher AUROC in four of the five equipment categories.

Table 3.3. Category-level comparison with Multi-Frequency on InsPLAD-fault

CategoryMulti-FrequencyWBGFlow
Glass Insulator87.7085.82
Lightning Rod Suspension99.3099.57
Upper Shackle94.5095.63
Vari-grip94.3094.57
Yoke Suspension97.6098.06
Macro-average94.794.73 ± 0.25

The table reports image-level AUROC (%) on InsPLAD-fault. The Multi-Frequency results come from Table I of the original paper and retain its reported precision. WBGFlow's category-level results are means over five training seeds. The macro-average row reports the mean and sample standard deviation. Bold indicates the higher reported value for the category.

Based on the table, WBGFlow leads by 0.27, 1.13, 0.27, and 0.46 percentage points on Lightning Rod Suspension, Upper Shackle, Vari-grip, and Yoke Suspension, respectively. Multi-Frequency leads by 1.88 percentage points on Glass Insulator. The reported macro-average AUROCs are equal at one-decimal precision, and WBGFlow has higher reported AUROC in four of the five categories. We will report the comparison at that precision in the revision.

3. Comparison with CLAdapter

CLAdapter reports 98.29% AUROC under the InsPLAD-fault supervised fault classification protocol. WBGFlow uses an unsupervised anomaly detection protocol with only normal training samples. CLAdapter's published setup includes 5,108 training images and 6,417 test images, and it uses annotations to adapt the model during training. WBGFlow trains without fault images or their labels. Ranking CLAdapter's 98.29% against WBGFlow's 94.73% would require matched training supervision and evaluation data. We will clarify this protocol difference in the revision.

4. Performance on MVTec AD

WBGFlow achieves 95.78% image-level AUROC on MVTec AD, 0.88 percentage points above the 94.90% reported in the original DifferNet paper. GLASS [1] and Bauer et al.'s Self-Supervised Autoencoders [2] report 99.9% and 99.8%, respectively, both higher than WBGFlow. We will include these published results in the revision and clarify that our primary application is anomaly detection in uncontrolled power-line inspection.

Reviewer_vUPQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes WBGFlow, a normalizing-flow-based framework for anomaly detection in UAV power-line inspection under varying environmental conditions. The method uses wavelet decomposition to construct three observation branches (RAW, LL, and HF), which are processed through shared grouped coupling layers with branch-identity conditioning. All three branches contribute to training, while only RAW and LL are used for anomaly scoring. Experiments on InsPLAD-fault and MVTec AD report image-level AUROCs of 94.62% and 95.78%, respectively. Ablation studies investigate the contributions of the proposed components and scoring strategy.

Critical strengths
  1. The paper addresses a practical problem in UAV-based power-line inspection, where complex backgrounds and varying imaging conditions make anomaly detection challenging.

  2. The separation of training and scoring branches is an interesting design choice. The method achieves improvements on InsPLAD-fault, with additional evaluation on MVTec AD.

Critical weaknesses
  1. The baseline comparison is incomplete. The paper mainly emphasizes improvements over DifferNet (2021), while omitting AttentDifferNet [1], which addresses the same task and reports a comparable AUROC of 94.34%. CLAdapter [2] also reports stronger results on InsPLAD-fault and provides publicly available code, although its experimental setting may differ.

  2. The novelty of the method needs further clarification. Wave-Flow[3] has already explored wavelet-based frequency information with normalizing flows for anomaly detection, including cross-frequency modeling. The authors should better explain how their approach differs from this work and what new contributions the branch-conditioned design provides.

[1] Vieira e Silva et al. Attention Modules Improve Image-Level Anomaly Detection for Industrial Inspection: A DifferNet Case Study. WACV, 2024.

[2] Li et al. Unleashing Foundation Vision Models: Adaptive Transfer for Diverse Data-Limited Scientific Domains. NeurIPS, 2025. Code: https://github.com/qklee-lz/CLAdapter

[3] Yang et al. Wavelet-Based Normalized Flow for Anomaly Detection in Photovoltaic Electroluminescent With Nonstationary Textures. IEEE Sensors Journal, 2025.

Questions for authors
  1. Are the baseline results in Table 1 independently reproduced or taken from previous publications? Several results match those reported in AttentDifferNet. Figures E.1–E.7 only compare WBGFlow with DifferNet. If the baselines were reproduced, could the authors provide qualitative comparisons with other representative methods? Otherwise, the sources of these results should be clarified, and relevant methods such as AttentDifferNet should be included in the comparison.

  2. CLAdapter is a recent open-source method evaluated on the same InsPLAD-fault dataset, reporting an AUROC of 98.29%. Could the authors evaluate this method and provide quantitative and qualitative comparisons? If the experimental protocols differ, the differences should be clearly explained.

  3. The separation of training and scoring branches appears to be a key contribution. Is there a theoretical or methodological reason for excluding HF during inference beyond the observed AUROC improvement? Moreover, RAW-only scoring slightly outperforms RAW+LL, making the choice of RAW+LL unclear. The relevant scoring ablations would be more informative in the main paper rather than only in the appendix.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 2 W1,W2

Response to Reviewer 2

Thank you for carefully reviewing our paper and recognizing the power-line inspection problem and our design separating training from scoring. We respond to each point below.

Critical Weakness 1

Thank you for pointing out the gaps in our comparison with related methods. We agree that AttentDifferNet is closely related to the task addressed in this paper and have added the results from its original paper. AttentDifferNet reports a macro-average image-level AUROC of 94.34% on InsPLAD-fault. We extended WBGFlow from two to five independent training seeds. Across the five runs, the macro-average AUROC is 94.73%, with a sample standard deviation of 0.25 percentage points. The mean is 0.39 percentage points higher than the published result for AttentDifferNet.

In addition to detection performance, we analyzed model parameter counts.

Table 2.1. Comparison of model parameter counts

MethodTotal parameters (M)Trainable parameters (M)
DifferNet174.559172.089
AttentDifferNet-SE∗174.572172.103
AttentDifferNet-CBAM∗174.573172.103
WBGFlow150.427147.958

The total count includes frozen and trainable parameters, with shared parameters counted once. We obtained the counts marked with an asterisk for AttentDifferNet-SE and AttentDifferNet-CBAM from implementations based on the architectures described in the original paper.

WBGFlow has 13.82% fewer total parameters than DifferNet. Both use the same frozen encoder, and the savings come from the flow module. Compared with DifferNet, WBGFlow reduces the total parameter count by 13.82% while achieving a higher detection AUROC. WBGFlow also has fewer parameters than the two AttentDifferNet variants implemented from the published architectures. These results show that WBGFlow improves on DifferNet while using fewer parameters.

We also examined CLAdapter's experimental protocol. Its reported 98.29% AUROC on InsPLAD-fault comes from supervised fault classification, whereas WBGFlow uses an unsupervised anomaly detection protocol with only normal training samples. The differences in supervision and data splits prevent a direct comparison under matched conditions.

In the revision, we will add the related methods and sources of their results and distinguish results obtained under different supervision protocols.

Critical Weakness 2

Thank you for pointing out the related work Wave-Flow. We agree that prior work has combined normalizing flows with frequency information extracted using wavelets.

WBGFlow's main contribution lies in shared density modeling across frequency observations conditioned on branch identity. Frequency decomposition provides the inputs for this modeling. Wave-Flow models correlations between frequency components. WBGFlow preserves the identity of the RAW, LL, and HF feature groups and uses shared coupling subnetworks to learn the normal feature distribution of each observation. Specifically, Grouped Flow performs invertible transformations within each branch and keeps the groups separate. Each coupling block uses a Branch Identity Embedding to identify the observation to the shared subnetworks, which then generate branch-specific affine transformation parameters. This design lets the three branches optimize shared parameters and use a single flow.

WBGFlow also uses frequency information differently during training and anomaly scoring. All three branches contribute to training, and the final anomaly score uses RAW and LL only. In the Upper Shackle ablation with the same RAW, LL, and HF inputs, Grouped Flow raises AUROC from 93.81% to 94.41%. Adding Branch Identity Embedding raises AUROC to 95.22%. With selective RAW+LL scoring, it reaches 96.14% (Table 3, configurations B–E). These comparisons directly measure the additional contribution of grouping and identity conditioning with the multi-frequency inputs held fixed. We will add Wave-Flow to the related work and clarify this difference in the methods section.

AuthorsOpenReview ↗

Response to Reviewer 2 Q1,Q2

Question 1

1. Sources of the results in Table 1

Table 1 includes results from the literature and from our own evaluations. For methods covered in the AttentDifferNet paper, we use its published results. For methods absent from that paper, we evaluated them under their respective configurations in the five-category unsupervised anomaly detection setting on InsPLAD-fault. Specifically, we evaluated methods including Dinomaly, INP-Former, ReContrast, and MSFlow under their respective experimental configurations. We use the results in the original AttentDifferNet paper for AttentDifferNet and the other baselines covered there. The 90.75% RD++ result in that paper is for CBAM–ResNet-18. In the revision, we will identify the source of each row and include AttentDifferNet-SE's reported 94.34%.

2. Qualitative comparison

Figures E.1–E.7 of the original manuscript primarily compare WBGFlow with DifferNet. We chose DifferNet because it is the normalizing flow method on which WBGFlow is based. Comparing their responses on the same anomalous samples helps show the effects of frequency-branch modeling and branch identity conditioning. Following the reviewer's suggestion, we prepared FastFlow overlays for the same InsPLAD-fault examples shown alongside DifferNet and WBGFlow in the updated qualitative comparisons.

Question 2

Thank you for suggesting CLAdapter. We checked its protocol and confirmed that CLAdapter and WBGFlow use different learning tasks and training protocols. CLAdapter's 98.29% AUROC on InsPLAD-fault comes from a supervised fault classification setting, whereas WBGFlow uses an unsupervised anomaly detection setting. CLAdapter adapts its model to the downstream task using labeled training samples. The 98.29% result is for the ConvNeXt-B variant. Its published setup includes 5,108 training images and 6,417 test images. WBGFlow follows the InsPLAD-fault unsupervised anomaly detection protocol and uses only normal training samples, with no anomalous training images or fault labels. The available training information and learning objectives differ. A direct comparison would require matched supervision and evaluation data. For reference, WBGFlow reports 94.62% in the original two-seed evaluation and 94.73% in the expanded five-seed evaluation.

A direct quantitative ranking or a qualitative comparison presented as a fair evaluation of the same task could therefore be misleading. We will clarify the protocol difference in the revision and discuss CLAdapter as a related supervised fault recognition method. The main comparison will cover methods evaluated under the same unsupervised anomaly detection protocol.

AuthorsOpenReview ↗

Response to Reviewer 2 Q3

Question 3

Thank you for asking about the separation between training and anomaly scoring. We explain this design through the optimization of shared parameters and the complementary contributions of the scoring branches.

1. Rationale for using HF during training and excluding it from scoring

WBGFlow's three frequency branches share the coupling subnetworks and retain separate latent representations throughout the invertible transformations.

During training, the three branches jointly optimize the shared parameters. The loss and its gradients are related as follows.

∇θLNF=13(∇θLRAW+∇θLLL+∇θLHF)\nabla_{\theta}\mathcal{L}_{\mathrm{NF}}=\frac{1}{3}(\nabla_{\theta}\mathcal{L}_{\mathrm{RAW}}+\nabla_{\theta}\mathcal{L}_{\mathrm{LL}}+\nabla_{\theta}\mathcal{L}_{\mathrm{HF}})

Although HF is excluded from the final anomaly score, its gradients update the shared parameters and affect the density transformations used by RAW and LL.

The high-frequency observations from HF can therefore contribute to shared modeling of normal features without requiring the HF anomaly score in the final ranking.

In inspection images, high-frequency information can indicate local defects and can also reflect background vegetation, power lines, or imaging noise. We use HF to support training and compute the final anomaly score from RAW and LL as follows.

Sfinal(x)=SRAW(x)+SLL(x)2S_{\mathrm{final}}(x)=\frac{S_{\mathrm{RAW}}(x)+S_{\mathrm{LL}}(x)}{2}

The shared-parameter optimization described above explains this design, and the ablation results support it.

On Upper Shackle, with RAW+LL scoring held fixed, joint training with three branches achieves 96.14% AUROC. Removing HF from training lowers AUROC to 94.30%, a difference of 1.84 percentage points.

In the five-category experiment at a fixed checkpoint, RAW+LL scoring achieves 94.320%, compared with 93.381% for RAW+LL+HF. Excluding the HF score improves AUROC in all five categories.

These experiments show the contribution of HF to joint training and the effect of excluding HF from the final score, respectively.

2. Why use RAW+LL instead of RAW-only?

Thank you for the suggestion. In the fixed-checkpoint experiment in Appendix B.3, the macro-average AUROC of RAW-only is 94.363%, 0.043 percentage points above the 94.320% of RAW+LL. To examine how the two scoring rules perform under changes in acquisition conditions, we compared RAW-only and RAW+LL on the same InsPLAD-fault test images.

Table 2.2. Comparison of RAW-only and RAW+LL on the original images and under seven controlled corruptions

Test conditionRAW-onlyRAW+LLFusion gain (pp)
Original images93.8693.97+0.11
Decreased brightness92.1592.49+0.34
Increased brightness91.4992.07+0.58
Decreased contrast92.2092.54+0.34
Increased contrast91.9091.82−0.08
Gamma change92.4092.71+0.31
Gaussian blur91.9592.91+0.96
Gaussian noise54.7364.15+9.42

The RAW-only and RAW+LL columns report the macro-average image-level AUROC (%) across the five InsPLAD-fault categories. The fusion gain is RAW+LL AUROC minus RAW-only AUROC, in percentage points (pp). Bold indicates the higher AUROC for each condition.

In this additional experiment, RAW+LL achieves 93.97% on the original images, compared with 93.86% for RAW-only. RAW+LL achieves higher AUROC under six of the seven corruptions. The gain under Gaussian blur is 0.96 percentage points, and AUROC under Gaussian noise rises from 54.73% to 64.15%. When contrast is increased, RAW-only is higher by 0.08 percentage points. These results support the use of fixed, equally weighted RAW+LL scoring. The low-frequency structure provided by LL improves anomaly ranking under most tested corruptions.

We will move the scoring ablations to the ablation section of the main paper. We will also add the comparison between RAW-only and RAW+LL on the original images and under seven controlled corruptions, so readers can see the effect of excluding the HF score and the benefits and trade-offs of adding LL under different test conditions.

Thank you for the detailed response and additional experiments, which address several of my concerns. However, the methodological novelty remains limited, and the improvement over AttentDifferNet is modest (0.39 percentage points). I consider this paper borderline and will maintain my original score. I encourage the AC to consider the perspectives of the other reviewers, particularly if they provide further feedback during the discussion.

Reviewer_ddLROpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper mainly proposed a method to fuse multi-spectral data input for image anomaly detection in the application of power line inspection. The proposed framework adapts normalizing flow to support inputs of RAW appearance, low-frequency structure, and high-frequency energy. The experiment shows SOTA averaged performance for a powerline anomaly detection dataset.

Critical strengths

Overall, the paper is easy to follow and clear. The experiment is extensive to ablate the modules and shows the proposed method can achieve SOTA performance on the targeted powerline inspection dataset.

Critical weaknesses

The proposed method only achieves a strong performance on the powerline inspection dataset but notably worse than other normalizing flow methods and many other previous SOTAs on the MVTec dataset as shown in Table 2.

Questions for authors
  1. Could you please elaborate more or make more analyses on the design choices of the ways to fuse different spectral inputs, ie, raw image, high-freq, and low-freq data? It would be better to understand the trade-off of different frequency of info for the powerline anomaly detection.

  2. Have you considered if the proposed multi-spectral decomposition modeling idea can be appliable to other normalizing flow architectures to further achieve stronger or efficient performance? I think the main concern of me on this paper is that it does not clearly analyze how different wavelets affect the paper and how to fully use it to strengthen the whole anomaly detection or localization method performance.

3 comments
AuthorsOpenReview ↗

Response to Reviewer 1 W1

Thank you for carefully reviewing our paper and recognizing WBGFlow's experimental results on anomaly detection in power-line inspection. We respond to each weakness and question below.

Weakness 1

WBGFlow achieves an image-level AUROC of 95.78% on MVTec AD, 0.88 percentage points above the 94.90% reported in the original DifferNet paper. This shows that our frequency-branch-conditioned modeling is effective on a conventional industrial anomaly detection dataset as well as on InsPLAD-fault. WBGFlow primarily addresses anomaly detection in uncontrolled power-line inspection. Compared with images from conventional industrial inspection settings, UAV inspection images typically contain more complex backgrounds and substantial variation in illumination and viewpoint. We study how shared conditional density modeling across frequency observations can improve anomaly detection under these conditions. On InsPLAD-fault, we extended the evaluation of WBGFlow from two to five independent training seeds. The five runs yield a macro-average image-level AUROC of 94.73%, with a sample standard deviation of 0.25 percentage points across seeds. The five-seed mean exceeds DifferNet's 92.46% by 2.27 percentage points, and the mean AUROC is higher than DifferNet's in all five equipment categories. These results demonstrate the effectiveness of the proposed method in its target application.

To examine how these results relate to uncontrolled acquisition conditions, we added a controlled robustness experiment. We applied illumination changes, contrast degradation, Gaussian blur, and noise to the same test images and compared WBGFlow with DifferNet.

Table 1.1. Detection performance under controlled corruptions on InsPLAD-fault

Test conditionDifferNetWBGFlowDifferNet dropWBGFlow drop
Clean89.6093.970.000.00
Brightness change(mean dark/light, severity2)86.2992.283.311.69
Contrast reduction (0.70)83.7792.545.831.43
Gaussian blur (radius1.0)85.4592.914.151.06
Gaussian noise (std0.03)58.5064.1531.1029.82
Mean over corruptions78.5085.4711.108.50

The DifferNet and WBGFlow columns report equally weighted macro-averages of image-level AUROC (%) across the five InsPLAD-fault equipment categories. The brightness row averages darkening and brightening at severity level 2. Drop is the difference between a model's clean AUROC and its AUROC under the corresponding corruption, in percentage points (pp). The last row averages the four corruptions with equal weight and excludes the clean condition. Values are reported to two decimal places.

WBGFlow achieves higher AUROC than DifferNet under all four controlled corruptions. It also shows a smaller drop from its own clean AUROC in each case. Across the four corruptions, WBGFlow averages 85.47% AUROC, 6.97 percentage points above DifferNet's 78.50%. The average drop decreases by 2.60 percentage points, from 11.10 for DifferNet to 8.50 for WBGFlow. The differences are most pronounced under contrast reduction and Gaussian blur. WBGFlow drops by 1.43 and 1.06 percentage points, respectively, compared with 5.83 and 4.15 for DifferNet. These results show that WBGFlow better maintains its detection performance under the tested imaging corruptions.

AuthorsOpenReview ↗

Response to Reviewer 1 Q1

Question 1

Thank you for raising this question. WBGFlow uses three branches, RAW, LL, and HF, to exploit complementary information from different frequency observations. Shared parameters support density modeling conditioned on branch identity. The design keeps the frequency observations in separate groups instead of concatenating their features.

RAW retains the color, texture, and structure of the original RGB image, preserving its full appearance. LL is the low-frequency approximation obtained by Haar wavelet decomposition and mainly captures coarse equipment structure. HF is the mean of the squared coefficients from the three directional high-frequency subbands, LH, HL, and HH, and represents edge and fine-detail energy.

The three branches extract features with a shared, frozen AlexNet and apply global average pooling separately. Grouped Flow then preserves the feature groups of the different branches. In each coupling block, a Branch Identity Embedding identifies the branch to the shared subnetworks, enabling them to generate affine transformation parameters for each observation. The three branches are thus optimized jointly through shared flow parameters. The feature groups are transformed separately.

During training, RAW, LL, and HF all contribute to maximum-likelihood optimization.

LNF=13(LRAW+LLL+LHF)\mathcal{L}_{\mathrm{NF}}=\frac{1}{3}(\mathcal{L}_{\mathrm{RAW}}+\mathcal{L}_{\mathrm{LL}}+\mathcal{L}_{\mathrm{HF}})

Because all three branches share the coupling subnetworks, gradients from the HF branch also update the parameters used by RAW and LL.

At inference, we compute the final anomaly score as follows.

Sfinal(x)=SRAW(x)+SLL(x)2S_{\mathrm{final}}(x)=\frac{S_{\mathrm{RAW}}(x)+S_{\mathrm{LL}}(x)}{2}

This design lets different frequency observations contribute differently to learning the normal distribution and to ranking anomalies. HF captures fine details and can also respond to high-frequency textures in complex backgrounds. We use HF to help train the shared density model and use RAW and LL to compute the final anomaly score.

The experiments support both design choices. First, HF benefits joint training. On Upper Shackle, training with all three branches yields 96.14% AUROC. Removing HF from training while keeping RAW+LL scoring reduces AUROC to 94.30%, a difference of 1.84 percentage points. Second, excluding the HF score improves anomaly ranking. In the five-category experiment at a fixed checkpoint, RAW+LL scores 94.320% macro-average AUROC, compared with 93.381% for RAW+LL+HF, with improvements in all five categories. These ablations evaluate the contribution of HF to training and the effect of selective scoring, respectively.

AuthorsOpenReview ↗

Response to Reviewer 1 Q2

Question 2

Thank you for the suggestion regarding the method's applicability. We discuss the choice of wavelet basis and the application to other flow architectures separately.

1. Effect of different wavelet bases

We use the Haar wavelet because its simple local computation efficiently extracts a low-frequency approximation and directional high-frequency details. The core design is independent of the specific Haar filter coefficients. The Branch-Conditioned Grouped Flow operates on the feature representations of the different observations, so its design can use other wavelet bases to generate low- and high-frequency observations. To evaluate the effect of the wavelet basis on detection performance, we compared Haar, db2, and sym4 while keeping the other architectural choices, training settings, and evaluation protocol fixed.

Table 1.2. Detection performance with different wavelet bases on InsPLAD-fault

Wavelet basisGlass Ins.Light RSUpper Sha.Vari-gripYoke Sus.Macro-average
Haar84.3699.5494.8294.7698.1294.32
db283.8399.3794.4194.4097.5893.92
sym483.4699.2594.3194.3397.8093.83

Results for each category are image-level AUROC (%). The macro-average weights the five equipment categories equally. Bold indicates the highest value in each column.

Haar achieves a macro-average AUROC of 94.32%, exceeding db2 and sym4 by 0.40 and 0.49 percentage points, respectively. It also achieves the highest AUROC in all five categories. The macro-averages of the three wavelet bases differ by at most 0.49 percentage points, indicating that the choice of basis had a small effect on overall detection performance in this experiment. We use Haar in the main model given its performance across the five categories and its simple local computation.

2. Applying multi-frequency decomposition to other normalizing flow architectures

Thank you for the suggestion. Wavelet decomposition takes place before feature extraction, so RAW, LL, and HF can serve as input observations for other flow architectures. We integrated the three observations into CFLOW and FastFlow and compared the original and multi-frequency inputs within each architecture. With each input setting, we tested fusion of all three branches and scoring with the third branch omitted to examine how frequency decomposition and score selection jointly affect detection performance and inference time.

Table 1.3. Applying multi-frequency decomposition to other normalizing flow architectures

Target flow architectureInput observationsFinal scoring methodMacro-average AUROC (%)Inference time (ms/original image)
CFLOWORIGINALFusion of all three branches82.49559.93
CFLOWORIGINALOmit the third branch83.14559.93
CFLOWRAW, LL, HFFusion of all three branches76.73451.94
CFLOWRAW, LL, HFOmit the third branch83.97451.94
FastFlowORIGINALFusion of all three branches79.56294.99
FastFlowORIGINALOmit the third branch78.64294.99
FastFlowRAW, LL, HFFusion of all three branches77.55302.62
FastFlowRAW, LL, HFOmit the third branch77.94302.62

AUROC is the equally weighted image-level macro-average across the five InsPLAD-fault categories. Original uses three resized inputs, S1 (448 × 448), S2 (224 × 224), and S3 (112 × 112). All three branches participate in training. Two-branch fusion omits S3 in Original and HF in the wavelet setting. For each configuration, it reuses the checkpoint selected with three-branch scoring. Inference time is measured per original image.

On CFLOW, multi-frequency inputs combined with RAW+LL scoring improve both detection performance and inference efficiency. With the third branch omitted in both settings, multi-frequency inputs raise macro-average AUROC from 83.14% to 83.97%, a gain of 0.83 percentage points, and reduce inference time from 559.93 to 451.94 ms per original image, a reduction of 19.29%. With multi-frequency inputs held fixed, excluding the HF score increases AUROC from 76.73% to 83.97%, a gain of 7.24 percentage points. The benefit of frequency decomposition therefore depends on how the branch scores are used in the final anomaly ranking.

On FastFlow, excluding the HF score with multi-frequency inputs also raises AUROC, from 77.55% to 77.94%. The best original-input result is 79.56% at 294.99 ms per image, compared with 77.94% at 302.62 ms for RAW+LL scoring. These experiments show that both flows can use multi-frequency inputs. In CFLOW, the input setting changes latency, while excluding the HF score improves AUROC at the same latency. We will add Table 3 and the analysis above to the revised manuscript.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.