A Hierarchical Transformation-Discriminating Generative Model for Few Shot Anomaly Detection

• The task is motivated by the scarcity of training images in visual domains and the ability of people to identify anomalies from few examples.Source: §1 · p. 1
• A hierarchical generative model captures internal patch statistics from one or a few images at multiple scales.Source: §1 · p. 1
• Conditioning the generator on image index extends single-image modeling to multiple training images and enables information sharing between them.Source: §1 · p. 1
• Multi-scale patch modeling covers both global properties such as color or structure and anomalies confined to localized regions.Source: §1 · p. 1
• At each scale, transformed generated patches are judged by a discriminator for both real-versus-fake status and transformation identity.Source: §1 · p. 1
• Across one-, five-, and ten-shot benchmarks, the method outperforms recent baselines, improves with more training samples, and better localizes anomalous patches in few-shot defect detection.Source: §1 · p. 1

• Reconstruction-based methods use low error for normal data and high reconstruction error for anomalies.Source: §2 · pp. 1–2
• Classification-based approaches include self-supervised proxy tasks, while this method places transformation classification within a hierarchical patch-based generative setting.Source: §2 · p. 2
• Generative distribution modeling includes autoencoders, variational autoencoders, and GANs, but a generative model alone may not suffice for out-of-distribution detection.Source: §2 · p. 2
• The approach combines a hierarchical model of multi-scale patch distributions with scale-dependent discrimination of patch realism and transformations.Source: §2 · p. 2
• Prior internal-statistics GANs model one image for generation, whereas this method targets anomaly detection from one or a few normal images.Source: §2 · p. 2

3. Method

• The task is to learn a normal-versus-anomalous classifier from a small i.i.d. sample of images drawn from the normal class.The setting includes k = 1 as well as other small training-set sizes.Source: §3 · p. 2

3.1. One Shot Anomaly Detection

• At every scale, discriminators classify transformed patches by transformation and distinguish fake patches; low correct-transformation votes across patches indicate anomaly, and the authors report improvement from the hierarchical model.The anomaly score aggregates patch votes across scales and transformations.Source: §3.1 · pp. 3–4

3.2. From one shot to few shot anomaly detection

• For few-shot training, one generator conditioned on the training-image index models each image distribution, avoiding separate models and enabling information sharing between images.The index is concatenated as an additional input channel.Source: §3.2 · p. 4

3.3. Defect Detection

• For defect detection, the method averages the 5% of patches with the lowest anomaly scores rather than averaging all patches.The selected patches are defined separately for each scale and transformation.Source: §3.3 · pp. 4–5

3.4. Architecture and training details

• Generators and discriminators use five convolutional blocks, yielding an 11 × 11 generator receptive field; training uses Adam at 0.0005 with α = 100.The finest scale is set to at most 64 × 64 pixels.Source: §3.4 · p. 5

4. Experiments

• Images were resized to 64 × 64 and evaluated by mean AUC over ten random training trials; benchmarks included Paris, CIFAR10, FashionMNIST, MNIST, and MVTec.MVTec includes 10 object and 15 texture categories, with 70 defect types in test images.Source: §4 · p. 5, §4.1 · p. 5, §4.2 · p. 6

4.1. Anomaly detection

• Our method outperforms all baselines across Paris, CIFAR10, FashionMNIST, and MNIST in one-shot anomaly detection.Figure 2 presents average AUC for these one-shot comparisons.Source: §4.1 · p. 5, Figure 2 · p. 6
• In five- and ten-shot detection, our method outperforms baselines on all four datasets; on Paris, five samples raise AUC from 66.6% to 79.8%.On CIFAR10, it also improves with more samples and outperforms baselines at 50 and 80 shots, while its gap from GEOM decreases as k increases.Source: §4.1 · p. 5, Figure 2 · p. 6, Figure 4 · p. 7

4.2. Defect detection

• On MVTec, our method outperforms all baselines in one-, five-, and ten-shot defect detection.Ours1 uses the standard transformations; Ours2 uses only rotations for comparison with DifferNet.Source: §4.2 · p. 7, Figure 5 · p. 8
• For one-shot MVTec localization, our method captures defect regions accurately, whereas DifferNet captures smaller regions.Figure 6 shows original images with anomaly regions delineated in red alongside localization maps from both methods.Source: §4.2 · p. 7, Figure 6 · p. 8

4.3. Ablation analysis

• Ablations show that the generator, transformations, and patch hierarchy are all needed for best performance, with hierarchy particularly important.The table reports average CIFAR10 AUC with or without each component.Source: Table 1 · p. 7, §4.3 · p. 8
• GEOM trained on generated images is less effective than the proposed method but more effective than few-shot GEOM.The paper reports these comparisons alongside the component ablations.Source: §4.3 · p. 8
• Using 5% of patches is best for one-shot detection, while 10% is best for five- and ten-shot settings; results remain stable at low patch fractions.Figure 7 plots average MVTec AUC by percentage of patches used.Source: §4.3 · p. 8, Figure 7 · p. 8

5. Conclusions

• The model incorporates self-supervised transformation classification within its discriminators and adds image-index conditioning to train across multiple images.The authors describe both the integration and multi-image training as distinctive aspects of their approach.Source: §5 · p. 8
• Without further modifications, the method does not scale well to hundreds or thousands of training images, and training becomes more involved as image count increases.Source: §5 · p. 8
• For the small training sets it targets, the method demonstrates one-class classification capabilities described as surprising given the literature’s emphasis on modeling variability between training samples.Source: §5 · p. 9