PCSNet: Prototypical FSAD for Industrial Inspection

• PCSNet is a few-shot anomaly detection framework that combines prototype-guided feature adaptation with context-aware segmentation for precise pixel-level localization.
• It employs synthetic pseudo anomalies to compact the normal feature distribution and enhance the separation between normal and anomalous data.
• PCSNet demonstrates robust performance on benchmarks like MVTec, MPDD, and automotive inspections, achieving high AUROC scores even in challenging FSAD scenarios.

PCSNet, short for Prototypical Learning Guided Context-Aware Segmentation Network, is a few-shot anomaly detection (FSAD) framework for industrial inspection settings in which only a handful of normal samples are available for a target category and no real anomalies are used for standard training. It is designed to address the domain gap between pre-trained representations and target FSAD scenarios by combining prototype-guided feature adaptation with a context-aware segmentation branch for pixel-level localization. The architecture comprises a Prototypical Feature Adaption (PFA) sub-network and a Context-Aware Segmentation (CAS) sub-network; on MVTec and MPDD it reports 94.9% and 80.2% image-level AUROC in an 8-shot scenario, respectively, and it is also evaluated on a real-world automotive plastic part inspection dataset (Jiang et al., 17 Dec 2025).

1. Problem setting and design rationale

PCSNet is formulated for the FSAD regime in which the training set contains only a few normal samples of a target class, while the test set contains both normal and anomalous images. The paper’s central premise is that direct transfer from a source-domain backbone, such as an ImageNet-pretrained WideResNet50, is insufficient because the target normal samples may remain scattered in feature space rather than forming a compact cluster. In the few-shot setting, that scatter is especially problematic because the available normal data are too limited to characterize the target distribution reliably (Jiang et al., 17 Dec 2025).

The method therefore pursues two explicit objectives. First, it seeks to compact the normal feature distribution in the target domain. Second, it seeks to increase separation between normal and anomalous features, even though real anomaly labels are unavailable during standard training. To supply negative supervision, PCSNet introduces synthetic anomalies—called pseudo anomalies—generated from training images. This design places feature adaptation and anomaly localization within the same framework rather than treating them as separate stages.

A plausible implication is that PCSNet is best understood not as a pure reconstruction-based anomaly detector, nor as a pure prototype classifier, but as a joint representation-learning and dense prediction system whose few-shot behavior depends on both compactness in embedding space and pixel-wise discrimination.

2. Architectural composition

PCSNet consists of two coupled sub-networks: PFA and CAS (Jiang et al., 17 Dec 2025).

The PFA sub-network contains a pre-trained feature extractor $E$, with WideResNet50 as the default backbone, together with a lightweight feature adaptor $A$. The adaptor is composed of three $1 \times 1$ convolution layers and one CoordConv layer. Its inputs are normal images $x_n$ from $D_{\text{train}}$ and corresponding synthetic abnormal images $x^a$, generated by the NSA method by cutting patches from training images and pasting them using Poisson image editing. The role of PFA is to produce adapted features whose geometry is more appropriate for the target category than the original pre-trained representation.

The CAS sub-network is an FPN-like segmentation head $S$. It takes as input the adapted features $F$, prototype-based similarity maps $S^{p\!y}(F)$, and multi-scale contextual features, and produces a pixel-wise anomaly score map. The paper frames CAS as the component responsible for fine-grained anomaly localization, especially once the feature space has been reshaped by prototype-guided adaptation.

The two branches are functionally complementary. PFA operates at the representation level, enforcing compactness for normal data and separation from synthetic abnormal data. CAS operates at the pixel level, converting adapted features and similarity cues into a dense anomaly map. This suggests that PCSNet distributes the FSAD problem across two coupled spaces: a prototype-centered feature space and a context-aware segmentation space.

3. Prototype-guided learning and optimization objectives

The total training objective is written as

$$L_{\text{total}} = L_{\text{NFC}} + L_{\text{AFS}} + \lambda_1 L_{\text{PDC}} + \lambda_2 L_{\text{SEG}},$$

with experimental settings

$A$0

The first term, Normal Features Compacting loss $A$1, constrains normal feature patches to remain within a hypersphere centered at prototypical features:

$A$2

Here, $A$3 is the number of feature patches, $A$4 is Euclidean distance, and $A$5 is the compactness radius. Patches inside the radius incur no penalty; patches outside are pushed toward the prototype region.

The second term, Abnormal Features Separation loss $A$6, pushes synthetic abnormal patches outside an enlarged boundary:

$A$7

where $A$8 is a radius relaxation coefficient. This loss does not force pseudo anomalies arbitrarily far from the normal region; it requires them to lie beyond the margin $A$9.

The third term, Pixel-level Disparity Classification loss $1 \times 1$0, is introduced for subtle anomalies that may remain close to normal features globally. PCSNet computes a disparity map

$1 \times 1$1

selects the top-$1 \times 1$2 pixels with largest disparity values,

$1 \times 1$3

and averages those values to obtain an image-level abnormality score,

$1 \times 1$4

The classification term is then

$1 \times 1$5

with $1 \times 1$6 for anomaly and $1 \times 1$7 for normal. The paper presents this mechanism as a way to amplify hard anomalies by focusing optimization on the most deviant local regions.

The fourth term, Segmentation loss $1 \times 1$8, supervises the CAS sub-network using masks derived from synthetic anomalies:

$1 \times 1$9

The paper describes this as a simple cross-entropy-style segmentation objective, with $x_n$0 and $x_n$1 denoting the predicted anomaly localization output and the ground-truth mask, respectively (Jiang et al., 17 Dec 2025).

Taken together, these losses encode a specific geometry: normals are pulled into a compact prototype-centered region, pseudo anomalies are excluded from that region, subtle discrepancies are emphasized through top-$x_n$2 disparity selection, and the resulting structure is projected into a pixel-level anomaly map.

4. Training and inference procedure

Training is joint over the PFA and CAS branches (Jiang et al., 17 Dec 2025). A batch of normal images is sampled from $x_n$3, and for each normal image a synthetic abnormal counterpart is generated using NSA. Both normal and pseudo-anomalous images are passed through the frozen extractor $x_n$4 and the learnable adaptor $x_n$5. Prototype features $x_n$6 are formed from normal patches, after which the model computes $x_n$7, $x_n$8, $x_n$9, and $D_{\text{train}}$0, and optimizes the total objective $D_{\text{train}}$1.

At inference time, the PFA-adapted feature extractor produces embeddings for the test image and the CAS branch outputs a pixel-wise anomaly score map. The paper applies Gaussian smoothing with $D_{\text{train}}$2 as post-processing. The resulting score map is used for both image-level anomaly scoring and pixel-level localization.

A central practical feature of PCSNet is its use of pseudo anomalies during training while remaining applicable to the standard few-shot normal-only scenario. The synthetic anomalies are not treated as realistic replacements for true defects; rather, they act as a supervisory proxy that sharpens the decision boundary and improves localization.

5. Datasets, shot settings, and reported performance

PCSNet is evaluated on MVTec AD, MPDD, and an in-house Automotive Plastic Part Inspection Dataset (APPD) (Jiang et al., 17 Dec 2025). MVTec AD contains 5354 high-resolution images, with 5 texture classes and 10 object classes; training is normal-only, and testing includes normal and anomalous images with pixel-level annotations. MPDD contains 6 classes of industrial painted metal parts under varying lighting, pose, distance, and background complexity. APPD contains 4082 normal training images, 1749 normal testing images, and 69 abnormal testing images, standardized to $D_{\text{train}}$3, and targets subtle defects including protrusion, spot, concave, and scratch.

The paper reports 2-shot, 4-shot, and 8-shot results on MVTec and MPDD, and few-shot results including up to 10-shot on APPD. The main reported metrics are image-level AUROC and pixel-level AUROC; for APPD it also reports AUPRO.

Dataset	Shot setting	Reported PCSNet performance
MVTec	2-shot	90.4 image-AUROC / 95.7 pixel-AUROC
MVTec	4-shot	92.1 / 96.5
MVTec	8-shot	94.9 / 97.1
MPDD	2-shot	67.7 image-AUROC / 92.7 pixel-AUROC
MPDD	4-shot	70.8 / 95.1
MPDD	8-shot	80.2 / 96.2
APPD	reported few-shot setting	92.8% anomaly detection accuracy / 92.8% AUPRO

On MVTec, the paper states that PCSNet is better than RD4AD, CFA, PatchCore, RegAD, RFR, and PACKD in most settings, and notes that at 8-shot the reported values are PCSNet: 94.9, PACKD: 95.3, RegAD: 91.2, and PatchCore: 94.3. On MPDD, the 8-shot image AUROC of 80.2% is reported as +9.7 over PACKD and +8.3 over RegAD. On the real-world APPD, PCSNet is reported to outperform RD4AD, CFA, and PatchCore substantially, especially under low-shot conditions (Jiang et al., 17 Dec 2025).

These results situate PCSNet as a method whose strongest reported gains arise on more difficult or more data-limited settings, particularly when subtle defects and target-domain mismatch make direct use of pre-trained features insufficient.

6. Ablations, limitations, and acronym disambiguation

The ablation study is central to the paper’s technical claims. Without the prototype-guided losses, performance drops substantially. Adding PDC yields large gains, which the paper interprets as evidence that pixel-level hard-sample mining is necessary for subtle anomalies. Adding NFC + AFS further improves performance by compacting normal features and separating anomalies, and the full combination performs best. The paper also reports that removing CAS causes many normal patches to receive high anomaly scores, increasing false positives, whereas adding CAS improves anomaly localization and reduces misclassification of normal regions; the reported gains are about +3.1% / +2.4% / +4.2% on MVTec across the shot settings. For CAS input design, concatenating extracted features $D_{\text{train}}$4 and similarity maps $D_{\text{train}}$5 performs better than using only $D_{\text{train}}$6. The method is reported to be relatively stable with respect to $D_{\text{train}}$7, with $D_{\text{train}}$8 giving the best balance. For pseudo-anomaly generation, NSA performs better than PII, FPI, and CutPaste, and the paper also notes that introducing a small number of real anomalies can improve performance further (Jiang et al., 17 Dec 2025).

The main limitation explicitly noted is that the information exchange between PFA and CAS remains relatively simple, mainly through concatenation of similarity maps and extracted features. Suggested future directions include more sophisticated fusion mechanisms and possibly attention-based interaction between the two sub-networks.

A common source of confusion is nomenclature. PCSNet in (Jiang et al., 17 Dec 2025) is distinct from several unrelated architectures with similar abbreviations: PCN-Miner for protein contact network analysis (Guzzi et al., 2022), PCN as a predictive coding network for object recognition (Han et al., 2018), Prototypical Clustering Networks (PCN) for dermatological diagnosis (Prabhu et al., 2018), PCNet as a pairwise comparison network for remote sensing scene classification (Yue et al., 2022), Particle Convolution Network (PCN) for jet physics (Shimmin, 2021), and Phase Collaborative Network (PCN) for two-phase medical image segmentation (Zheng et al., 2018). Within the FSAD literature represented here, PCSNet specifically denotes the prototype-guided anomaly detection and segmentation framework introduced for industrial inspection (Jiang et al., 17 Dec 2025).