Anomaly-Guided Self-Supervised Pretraining

Updated 4 July 2026

AGSSP is a self-supervised learning framework that uses anomaly priors to shape pretraining, making anomaly detection and localization more precise.
It integrates both synthetic anomaly cues and normal manifold reconstruction, combining multi-scale features with specialized loss functions.
Empirical results demonstrate improved metrics (e.g., AUROC, AP) across diverse applications such as industrial inspection, ECG analysis, and medical imaging.

Anomaly-Guided Self-Supervised Pretraining (AGSSP) denotes a class of self-supervised learning schemes in which the pretraining signal is explicitly shaped by anomaly priors rather than by generic invariances alone. In the cited literature, this guidance may be direct, as in forged defects, pseudo-anomaly patches, pasted contextual inconsistencies, or anomaly-map distillation, or indirect, as in reconstructing masked normal data so that departures from a learned normal manifold emerge as elevated residuals, uncertainty, or energy. Across industrial inspection, electrocardiography, hyperspectral imaging, medical image analysis, urban scene understanding, and underwater depth estimation, AGSSP recurrently couples anomaly-aware perturbation or scoring with a representation of normal structure, so that downstream inference becomes a deviation-from-normal problem with explicit localization rather than a purely discriminative classification task (2209.12440, Jiang et al., 2024, Huang et al., 20 Mar 2026, Liu et al., 2024, Tu et al., 2023, Ding et al., 2024).

1. Conceptual definition and scope

In the ECG formulation, AGSSP is a self-supervised anomaly detection stage that learns the manifold of normal cardiac physiology while modeling multi-scale signal structure and demographic or physiologic attributes; anomalies are not injected as labeled classes, but are expected to appear at test time as reconstruction error, uncertainty, and attribute inconsistency relative to the learned normal manifold (Jiang et al., 2024, Huang et al., 20 Mar 2026). In visual anomaly detection, by contrast, the guidance is often explicit: forged anomalies are synthesized in salient regions, local perturbations are blended into otherwise normal images, or pseudo-abnormal classes are created so that a backbone must separate normal and abnormal-like content during pretraining (2209.12440, Zou et al., 2022, Tian et al., 2021).

A second axis of variation is the source of the anomaly prior. Some methods derive it from synthetic corruptions or copy-paste operations on normal data, including polygonal prisms in hyperspectral cubes, SmoothBlend perturbations in industrial images, or Poisson-blended lesions in medical scans (Liu et al., 2024, Zou et al., 2022, Müller et al., 2023). Others use a task-induced prior: SLEEG defines anomaly likelihood through the residual of segmentation energy, while WaterMono derives anomaly masks from teacher photometric inconsistency under view synthesis (Tu et al., 2023, Ding et al., 2024). This breadth is important because AGSSP is not a single loss or architecture; it is a design principle in which self-supervision is intentionally biased toward anomaly-sensitive structure.

The literature also distinguishes AGSSP from purely reconstruction-based or purely contrastive pretraining. Reconstruction alone can reconstruct anomalies too well, whereas generic contrastive learning can emphasize global semantics while neglecting fine local deviations. AGSSP methods therefore tend to add one or more of the following: synthetic anomaly exposure, normal-prior extraction, uncertainty modeling, task-energy residuals, dynamic masking, or representation geometry that explicitly separates normal and abnormal regions (Bauer et al., 2022, Rafiee et al., 2022, Yao et al., 7 Nov 2025).

2. Core algorithmic motifs

A canonical image-domain instantiation is the Self-Supervised Guided Segmentation Framework (SGSF). Its Saliency Augmentation Module constructs forged anomalies in salient foreground regions by combining a saliency map, Perlin noise, and adaptive thresholding to obtain irregular masks $M$ , then synthesizes a training image

$x_f = A \odot (1-M) + \alpha(B \odot M) + (1-\alpha)(A \odot M).$

Its Self-supervised Guidance Network reconstructs similar normal images to produce multi-scale guidance features, and the segmentation module fuses those normal priors with encoder features of the current image. Training uses

$L_{total} = \lambda L_{Self} + L_{Seg},$

with focal loss for segmentation and reconstruction loss for the guidance branch, so anomaly localization is conditioned on what normal should look like rather than on forged defects alone (2209.12440).

A related but theoretically different lineage is the Projecting Autoencoder view of self-supervised anomaly training. Here the input is a partially corrupted normal image, and the model minimizes a masked reconstruction loss that weights corrupted and uncorrupted regions separately. Under the paper’s Markov Random Field assumptions, the learned mapping resembles a non-linear orthogonal projection onto the manifold of normal samples, which explains why reconstruction error can become a localization signal rather than a mere denoising artifact (Bauer et al., 2022). SPD modifies contrastive pretraining in another direction: weakly augmented views are positives, while locally perturbed same-image views generated by SmoothBlend are negatives, with

$\mathcal{L}_{SPD} = \cos(z_i,\tilde z_i^-) - \cos(z_i,\tilde z_i^+),$

thereby enforcing sensitivity to subtle local differences while keeping robustness to slight global changes (Zou et al., 2022).

Signal-domain AGSSP is exemplified by the rare-ECG framework. It uses masked global and local cross-restoration, trend-assisted restoration, and patient attribute prediction. The uncertainty-aware restoration loss is combined with trend and attribute losses as

$\mathcal{L}_{AD} = \mathcal{L}_{res} + \alpha \mathcal{L}_{trend} + \beta \mathcal{L}_{pred}, \quad \alpha=\beta=1.0,$

and inference produces a pointwise anomaly map

$S(x_{test}) = S_g(x_{test}) + S_l(x_{test}),$

with record-level score

$A(x_{test}) = \frac{1}{D}\sum_{k=1}^{D} S(x_{test})^k.$

The guidance is indirect: the model learns normal rhythm and morphology so that rare deviations become measurable residuals (Huang et al., 20 Mar 2026).

Energy- and teacher-guided forms of AGSSP replace synthetic anomaly labels with task-derived inconsistency. SLEEG computes segmentation energy

$E_\theta(x,i) = -\log \sum_k \exp(z_{\theta,k}(x,i)),$

then trains a task-agnostic binary estimator and a task-oriented residual estimator, the latter enforcing margin-based separation between anomaly and normal pixels inside pasted regions (Tu et al., 2023). WaterMono uses a frozen teacher to synthesize neighboring views, computes a blurred photometric error, thresholds it by an exponential moving percentile to form a Teacher-Guided Anomaly Mask, and then trains a student with masked photometric loss, selective 3D-consistency distillation, and rotated distillation for roll robustness (Ding et al., 2024).

3. Representative instantiations across domains

The following examples illustrate how AGSSP is operationalized in different modalities.

Domain and method	AGSSP mechanism	Reported highlight
Industrial anomaly segmentation, SGSF (2209.12440)	Saliency-guided forged anomalies plus self-supervised normal-prior fusion	MVTec AUROC_det 96.7%, AUROC_loc 97.2%, AP_loc 73.8%
Rare ECG diagnosis (Huang et al., 20 Mar 2026)	Masked global/local restoration, trend-assisted restoration, attribute prediction	rare AUROC 94.7%; common–rare AUROC gap reduced by 73%
Metallic surface defect detection (Liu et al., 23 Sep 2025)	Anomaly-map-guided backbone pretraining plus pseudo-defect-box detector pretraining	up to 10% mAP@0.5 and 11.4% [email protected]:0.95 over ImageNet-based models
Hyperspectral anomaly detection, SAP (Liu et al., 2024)	Pseudo-anomaly prisms learned as anomaly prior within an LRR framework	average AUC(Pd,Pf) 0.99365
Urban scene anomaly segmentation, SLEEG (Tu et al., 2023)	Energy-guided likelihood estimation and adaptive refinement of pasted anomaly pixels	comparable to supervised competitors without auxiliary data or synthetic models
Underwater monocular depth, WaterMono (Ding et al., 2024)	Teacher-guided anomaly masking, enhancement-boosted photometrics, rotated distillation	Abs Rel 0.099 vs Lite-Mono 0.116

Medical-image AGSSP adds two further patterns. P-PII injects smoothly blended synthetic anomalies through probabilistic Poisson image interpolation, couples them with probabilistic backbones such as MC Dropout U-Net and PHiSeg, and reports gains up to +18% pixel AUROC and +45% sample AUROC on JSRT, with up to +53% pixel AUROC on LIDC-IDRI for PHiSeg relative to its baseline (Müller et al., 2023). PMSACL instead defines one normal class and multiple pseudo-abnormal classes generated by MedMix and related strong augmentations, then enforces dense clustering around frozen class centers for downstream UAD; on Hyper-Kvasir, PMSACL-PaDiM reaches AUROC 0.996 (Tian et al., 2021).

This diversity shows that AGSSP is less a single algorithm than a reusable decomposition: anomaly prior construction, self-supervised objective, and anomaly-scoring rule are chosen to match the modality’s structure.

4. Representation geometry, losses, and architectural patterns

One recurring pattern is the coexistence of normal-prior extraction and anomaly-sensitive separation. SGSF reconstructs similar normal images to obtain decoder guidance features, whereas ECG AGSSP fuses global rhythm and local heartbeat morphology through self-attention and outputs both reconstructions and uncertainty maps (2209.12440, Huang et al., 20 Mar 2026). In both cases, anomaly localization is conditioned on multi-scale normal semantics instead of being inferred from low-level residuals alone.

A second pattern is explicit control of feature geometry. PMSACL computes class centers

$\mathbf{c}_n = \mathbb{E}[f_\theta(a(\mathbf{x}))],$

then combines a multi-centering loss, a temperature-scaled contrastive loss with reduced same-class repulsion, augmentation classification, and patch-location prediction. The objective creates compact within-class clusters for the normal class and pseudo-abnormal subclasses, rather than the sparse normal distributions often induced by standard InfoNCE-style training (Tian et al., 2021). ADPretrain, although supervised at the feature and pixel level on RealIAD, advances a closely related anomaly-centric geometry: residual features are matched against normal references, then separated by angle-oriented and norm-oriented contrastive losses. The accompanying synthesis extends these objectives into a label-free AGSSP variant by replacing real anomaly labels with pseudo anomalies while keeping the same center-shifted angle separation and norm polarization (Yao et al., 7 Nov 2025). A plausible implication is that AGSSP can be interpreted not only as data synthesis but also as controlled shaping of the latent space around a normal center.

Self-distillation variants take a different route. The DINO-style OOD framework trains a teacher and student on in-distribution positives while forcing uniform student outputs on anomaly-like negatives created by shifting transformations such as rotations. The total loss

$L_{total} = L_{pos} + \lambda L_{neg}$

does not require explicit anomaly labels, yet it still injects anomaly guidance by penalizing confident in-distribution assignments on semantically shifted examples (Rafiee et al., 2022). This is closely aligned with AGSSP’s general principle: anomaly cues need not be true anomalies, but they must perturb the representation in a way that is relevant to downstream rejection.

Loss design is correspondingly heterogeneous. Reconstruction-centric methods use weighted masked losses, uncertainty-normalized residuals, SSIM terms, or focal segmentation losses (Bauer et al., 2022, 2209.12440, Huang et al., 20 Mar 2026). Contrastive methods use InfoNCE-like objectives, center losses, or same-image local-negative terms (Tian et al., 2021, Zou et al., 2022). Task-guided methods employ pairwise hinge losses on energy residuals or correlation-based distillation on transformed pseudo labels (Tu et al., 2023, Ding et al., 2024). What unifies them is not mathematical form but the role of the loss: it must make anomaly sensitivity a pretraining objective rather than an accidental byproduct.

5. Empirical performance and comparative evidence

Image-domain anomaly localization results show that AGSSP often improves both detection and pixel precision. SGSF reports MVTec image-level AUROC 96.7%, pixel-level AUROC 97.2%, and AP_loc 73.8%, with AP_loc exceeding DRAEM by about 5% and PaDiM by about 18%; removing the self-supervised module drops AP_loc by 17%, and replacing saliency-aware Perlin with global Perlin degrades both detection and localization (2209.12440). The Projecting Autoencoder reports MVTec average pixel-level AUROC 99.2 and image-level AUROC 99.8 for Model II, supporting the claim that anomaly-guided masked reconstruction can compete with or surpass memory-bank and GAN-based approaches without external pretraining (Bauer et al., 2022).

Pretraining that explicitly targets local discrepancy or anomaly-like negatives also shows consistent transfer gains. SPD improves anomaly segmentation AU-PR by 5.9% over SimSiam and 6.8% over supervised pretraining in the 2-class high-shot regime on VisA, while also improving most 1-class PaDiM settings on VisA and MVTec (Zou et al., 2022). In unsupervised OOD detection, self-distillation with rotated auxiliary negatives reaches 94.20 AUROC for CIFAR-10 versus CIFAR-100 in the combined setting, and Rot. DTI yields average AUROC 94.89 compared with 89.47 for DINO without negative guidance (Rafiee et al., 2022). These results support the more general observation that preserving low-level statistics while perturbing higher-level semantics can be an effective anomaly proxy.

Large-scale pretraining on industrial data strengthens this point. The metallic-surface AGSSP framework reports up to a 10% improvement in [email protected] and 11.4% in [email protected]:0.95 relative to ImageNet-based models, with particularly large gains in data-scarce regimes such as Casting Billet-Mini50, where CSPDarknet improves from 23.6/7.9 to 55.8/29.5 and Swin-Base from 15.3/5.3 to 56.3/29.8 (Liu et al., 23 Sep 2025). This directly addresses the domain-gap argument that natural-image pretraining is poorly aligned with metallic textures and subtle defect boundaries.

Biomedical and physics-grounded results show a similar pattern. The demographic-aware ECG framework reports rare-subset AUROC 94.7%, sensitivity 92.2%, specificity 92.5%, detection AUROC 91.2%, localization AUROC 76.5%, and Dice 65.3, while also achieving PTB-XL AUROC 89.6% and Renji AUROC 85.4% under external evaluation (Huang et al., 20 Mar 2026). P-PII raises JSRT pixel-wise AUROC from 0.80 to 0.90 for U-Net and yields up to +53% pixel AUROC for PHiSeg on LIDC-IDRI (Müller et al., 2023). WaterMono reduces FLSea Abs Rel from 0.116 to 0.099 with the same 3.1M parameter count as Lite-Mono, and its full configuration improves over each individual component, including Teacher-Guided Anomaly Mask, image-enhancement boosting, selective distillation, and rotated distillation (Ding et al., 2024).

6. Limitations, misconceptions, and future directions

A common misconception is that AGSSP is equivalent to synthetic anomaly generation. The literature shows otherwise: some methods depend on synthetic defects or pseudo-abnormal classes, but others rely on energy residuals, teacher photometric failures, or normal-manifold restoration with attribute prediction (Tu et al., 2023, Ding et al., 2024, Huang et al., 20 Mar 2026). AGSSP is therefore better understood as anomaly-centered self-supervision, not merely anomaly synthesis.

Another misconception is that anomaly guidance eliminates the forged-to-real gap. Several papers explicitly note residual mismatch. SGSF observes that semantic missing-component anomalies, such as the “transistor” category, are not fully captured by saliency-guided forged defects; poor saliency maps can also produce degenerate masks (2209.12440). The metallic-surface framework notes that repetitive textures can still trigger false positives and extremely subtle defects can be missed, even with knowledge-enhanced anomaly maps and adaptive thresholds (Liu et al., 23 Sep 2025). SAP in hyperspectral detection similarly depends on the pseudo-anomaly generator; if its polygonal prisms do not span the diversity of real anomalies, the learned prior can be biased (Liu et al., 2024).

Fairness, calibration, and domain shift remain active concerns. The demographic-aware ECG study reports slightly lower AUROC and specificity for ages under 10 and over 90 and explicitly states that no fairness regularizers were used, only attribute prediction and subgroup evaluation (Huang et al., 20 Mar 2026). WaterMono notes that per-scene attenuation and per-image backscatter are estimated heuristically and that the two-stage teacher–student design adds overhead (Ding et al., 2024). These findings suggest that AGSSP often improves robustness without solving calibration or subgroup reliability automatically.

The forward trajectory in the cited work is toward hybridization. Proposed directions include incorporating real anomaly cues when available through semi-supervised curricula, explicitly aligning forged and real anomaly distributions, adding subgroup-aware optimization or demographic reweighting, learning adaptive thresholds for pseudo labels, refining pseudo-anomaly generators to better encode spatial or relational semantics, and unifying teacher–student stages through EMA-style self-distillation (2209.12440, Liu et al., 23 Sep 2025, Huang et al., 20 Mar 2026, Liu et al., 2024, Ding et al., 2024). This suggests that AGSSP is evolving from a collection of task-specific tricks into a broader methodology for task-aligned representation learning under anomaly scarcity.