Multi-Scale Gaussian Anomaly Generator

• MGAG is a multi-scale anomaly simulation method that injects zero-mean Gaussian noise at three key feature levels to create plausible abnormal samples.
• It perturbs the input image, aggregated feature map, and discriminator input to enforce robust, scale-aware learning in multimodal systems (RGB and depth).
• Experimental validations on industrial datasets show MGAG achieves superior anomaly detection and localization compared to single-scale techniques.

The Multi-Scale Gaussian Anomaly Generator (MGAG) is a noise-injection methodology designed to synthesize diverse, semantically meaningful anomalies at multiple levels of a deep neural network architecture. Developed in the context of BridgeNet, a unified multimodal framework for bridging 2D and 3D industrial anomaly detection, MGAG addresses the challenge of simulating plausible abnormal samples when true anomaly data are scarce—particularly in multimodal scenarios involving both RGB and depth information. By distributing Gaussian perturbations across distinct abstraction levels within the network, MGAG compels discriminators to learn more robust, scale-aware anomaly representations, surpassing the efficacy of single-scale or Perlin-noise-only baselines (Xiang et al., 25 Jul 2025).

1. Conceptual Overview

MGAG operates as an anomaly simulation module that generates synthetic anomalies by injecting zero-mean Gaussian noise at three strategically chosen points (“scales”) within the feature extraction and discrimination pipeline. These points correspond to: the raw input image (G1), the fused feature map prior to discrimination (G2), and the discriminator’s input (G3). During training, MGAG produces perturbed data at each scale, which are labeled as anomalous, while corresponding clean features are labeled as normal. The discriminator is consequently trained to differentiate between these classes using a binary cross entropy (BCE) loss.

MGAG is implemented alongside the Unified Texture Anomaly Generator (UTAG) within BridgeNet, leveraging the shared architecture to jointly enrich the diversity of simulated anomalies for both RGB and depth modalities. The resultant learning process enables end-to-end optimization for robust detection of real-world, multimodal anomalies (Xiang et al., 25 Jul 2025).

2. Mathematical Formulation

Let $i$ index the training sample and $M \in \{\mathrm{rgb}, \mathrm{depth}\}$. For a batch, three zero-mean Gaussian noise tensors are sampled:

• $$\epsilon_{\mathrm{FE},M,i} \sim \mathcal{N}(0, \sigma_1)$$ (image-level; G1)
• $$\epsilon_{\mathrm{FA},M,i} \sim \mathcal{N}(0, \sigma_2)$$ (feature map; G2)
• $$\epsilon_{\mathrm{D},i} \sim \mathcal{N}(0, \sigma_3)$$ (discriminator input; G3)

with $\sigma_1 > \sigma_2 > \sigma_3 > 0$; paper defaults are $\sigma_1 = 0.12, \sigma_2 = 0.06, \sigma_3 = 0.02$.

The injected noisy variants are:

• $$x_{M,i}^{-} = x_{M,i} + \epsilon_{\mathrm{FE},M,i}$$ (G1)
• $$s_{M,i}^{-} = s_{M,i} + \epsilon_{\mathrm{FA},M,i}$$ (G2)
• $d_i^{-} = d_i + \epsilon_{\mathrm{D},i}$ (G3)

Here, $M \in \{\mathrm{rgb}, \mathrm{depth}\}$0 denotes the input image, $M \in \{\mathrm{rgb}, \mathrm{depth}\}$1 the fused and aggregated feature map, and $M \in \{\mathrm{rgb}, \mathrm{depth}\}$2 the input to the discriminator.

3. Algorithmic Steps and Scale-Wise Noise Injection

The MGAG operates within the following pipeline:

1. Feature Extraction: RGB and depth images, $M \in \{\mathrm{rgb}, \mathrm{depth}\}$3 and $M \in \{\mathrm{rgb}, \mathrm{depth}\}$4, are processed by the shared backbone $M \in \{\mathrm{rgb}, \mathrm{depth}\}$5 to obtain multi-layer features.
2. Feature Aggregation: Layer outputs are aggregated, upsampled, and concatenated per modality to obtain $M \in \{\mathrm{rgb}, \mathrm{depth}\}$6.
3. Feature Fusion and Discrimination: Concatenated modality features are passed through a fusion adaptor $M \in \{\mathrm{rgb}, \mathrm{depth}\}$7 (bias-free, fully-connected, no batch-norm), yielding $M \in \{\mathrm{rgb}, \mathrm{depth}\}$8 for each sample. These are then input to the discriminator $M \in \{\mathrm{rgb}, \mathrm{depth}\}$9.
4. Noise Injection: Independent Gaussian noise is sampled and injected at G1, G2, G3; perturbed samples are optionally further processed using “selective modality,” meaning that in each minibatch either one or both modalities may be perturbed.
5. Discriminator Training: Clean features are labeled negative (normal), noisy features positive (anomalous), and the BCE loss is computed to train $$\epsilon_{\mathrm{FE},M,i} \sim \mathcal{N}(0, \sigma_1)$$0.

The following pseudocode, paraphrased to align with the notation and processes described above, summarizes the procedure:

$$\epsilon_{\mathrm{FA},M,i} \sim \mathcal{N}(0, \sigma_2)$$9

4. Implementation Architecture and Training

Relevant architectural and training specifications are summarized below.

Component	Specification	Notes
Noise Scales	G1, G2, G3 (in order)	3 levels
Noise Stddev	$$\epsilon_{\mathrm{FE},M,i} \sim \mathcal{N}(0, \sigma_1)$$1, $$\epsilon_{\mathrm{FE},M,i} \sim \mathcal{N}(0, \sigma_1)$$2, $$\epsilon_{\mathrm{FE},M,i} \sim \mathcal{N}(0, \sigma_1)$$3	$$\epsilon_{\mathrm{FE},M,i} \sim \mathcal{N}(0, \sigma_1)$$4
Fusion Adaptor $$\epsilon_{\mathrm{FE},M,i} \sim \mathcal{N}(0, \sigma_1)$$5	Linear, no bias, no batch-norm, $$\epsilon_{\mathrm{FE},M,i} \sim \mathcal{N}(0, \sigma_1)$$6	Shared weights
Discriminator $$\epsilon_{\mathrm{FE},M,i} \sim \mathcal{N}(0, \sigma_1)$$7	Linear $$\epsilon_{\mathrm{FE},M,i} \sim \mathcal{N}(0, \sigma_1)$$8 BatchNorm $$\epsilon_{\mathrm{FE},M,i} \sim \mathcal{N}(0, \sigma_1)$$9 LeakyReLU(0.3) $$\epsilon_{\mathrm{FA},M,i} \sim \mathcal{N}(0, \sigma_2)$$0 Linear $$\epsilon_{\mathrm{FA},M,i} \sim \mathcal{N}(0, \sigma_2)$$1 Sigmoid	Input/hidden dim $$\epsilon_{\mathrm{FA},M,i} \sim \mathcal{N}(0, \sigma_2)$$2 (e.g. 1536)
Optimizer	AdamW	Batch size 4, 160 epochs
Learning Rate	$$\epsilon_{\mathrm{FA},M,i} \sim \mathcal{N}(0, \sigma_2)$$3 ($$\epsilon_{\mathrm{FA},M,i} \sim \mathcal{N}(0, \sigma_2)$$4, $$\epsilon_{\mathrm{FA},M,i} \sim \mathcal{N}(0, \sigma_2)$$5), $$\epsilon_{\mathrm{FA},M,i} \sim \mathcal{N}(0, \sigma_2)$$6 ($$\epsilon_{\mathrm{FA},M,i} \sim \mathcal{N}(0, \sigma_2)$$7)
Training protocol	Clean: label 0; Noisy: label 1; BCE loss	Selective Modality enabled

Selective modality, realized by randomly zeroing out one or both modalities per minibatch, constrains the model from overfitting to anomalies affecting both RGB and depth simultaneously.

5. Comparative Analysis with Prior Methods

MGAG extends prior anomaly generators—such as Perlin noise-based simulators—by distributing Gaussian perturbations not just at the input or single feature level, but across three semantic scales. Prior simulators operate at a single point: either by applying Perlin noise to the raw point cloud or by injecting disturbances at an intermediate feature level. MGAG applies large perturbations at the image scale (G1) and progressively smaller perturbations at deeper semantic representations (G2, G3). This scale hierarchy compels the downstream discriminator to recognize anomalies arising from both low-level color/depth shifts and high-level representational distortions.

An additional mechanism, “Selective Modality,” mitigates the risk that the discriminator simply detects dual-modal anomalies by introducing perturbations in only one modality (RGB or depth) at random during training epochs.

6. Experimental Validation and Metrics

MGAG has been extensively validated on the MVTec-3D AD dataset (10 classes; 2,656 train normal, 1,197 test) and the Eyecandies dataset (6 classes, RGB + depth). Performance is evaluated using image-level AUROC (I-AUROC), pixel-level AUROC (P-AUROC), and per-region overlap (P-AUPRO). Results are summarized below (Xiang et al., 25 Jul 2025):

Configuration	I-AUROC	P-AUPRO
UTAG only	0.946	0.949
MGAG only	0.969	0.965
UTAG + MGAG (no selection)	0.992	0.975
UTAG + MGAG (with selection)	0.993	0.977
Prior SOTA (3D+RGB)	0.978	0.972
BridgeNet (MGAG+UTAG+selection)	0.993	0.977

Ablation studies demonstrate:

• Multi-scale noise with $$\epsilon_{\mathrm{FA},M,i} \sim \mathcal{N}(0, \sigma_2)$$8 outperforms equal-σ configurations.
• Layer-wise injection at G2+G3 is strong; best results with G1+G2+G3.
• The UTAG+MGAG combination, especially with selective modality, achieves the highest reported performance.

7. Impact and Applications

MGAG advances the field of industrial anomaly detection by enabling more robust simulation of plausible defects across both 2D appearance and 3D structure. This multi-scale framework:

• Produces more diverse and challenging training anomalies, enhancing model generalizability.
• Enables unified architecture for RGB and depth, obviating the need for complex multimodal feature fusion.
• Demonstrably outperforms prior single-scale or Perlin-noise approaches in standard industrial AD benchmarks, with systematic gains in both detection and localization accuracy.

A plausible implication is that multi-scale perturbation generators such as MGAG may generalize to other modalities and domains where anomaly data are rare, provided a shared feature backbone and fusion strategy are available. This suggests potential future applications in medical imaging, remote sensing, and security, where robust, scale-aware anomaly detection remains a central challenge (Xiang et al., 25 Jul 2025).