Visual Anomaly Detection

Module: Computer Vision | Difficulty: Intermediate

Anomaly Score

where is a pretrained feature extractor and are normal reference features.

PatchCore

Uses coreset subsampling of nearest-neighbor features.

Feature Embedding Distance

Student-Teacher Distillation

import torch
import torch.nn as nn
import torchvision.models as models

class PatchCore:
    def __init__(self, backbone='resnet18', num_features=256):
        model = models.__dict__[backbone](pretrained=True)
        self.feature_extractor = nn.Sequential(*list(model.children())[:-2]).eval()
        self.memory_bank = []
    
    def fit(self, normal_loader):
        for images in normal_loader:
            with torch.no_grad():
                feats = self.feature_extractor(images)
                B, C, H, W = feats.shape
                feats = feats.permute(0, 2, 3, 1).reshape(-1, C)
                self.memory_bank.append(feats)
        self.memory_bank = torch.cat(self.memory_bank)
    
    def score(self, images):
        with torch.no_grad():
            feats = self.feature_extractor(images)
            B, C, H, W = feats.shape
            feats = feats.permute(0, 2, 3, 1).reshape(-1, C)
            dists = torch.cdist(feats, self.memory_bank)
            min_dists, _ = dists.min(dim=1)
            scores = min_dists.reshape(B, H, W)
        return scores

Key Takeaways

• Unsupervised anomaly detection requires only normal training data
• PatchCore achieves strong performance with nearest-neighbor search
• Industrial inspection demands low false-positive rates