A digital image is visual data represented through pixel values or derived visual features. In outlier detection, the main question is usually whether the depicted object belongs to a type present in training, rather than whether its colors, angle, shading, or background look unusual. section 16.3
Image-based outlier detection commonly uses several complementary signals:
- Reconstructive models, such as autoencoders, compress images and use reconstruction error as an outlier score. section 16.3.1
- Generative models, including variational autoencoders and generative adversarial networks, model or regenerate image data. section 16.3.1
- Confidence-based methods treat uniformly low classifier probabilities as evidence that an image may be out of distribution. section 16.3.1
- Feature modeling applies standard detectors such as isolation forests or k-nearest neighbors to object embeddings. section 16.3.1
- Perturbation tests rotate, discolor, crop, or stretch images and check whether the model predicts the applied transformation reliably. section 16.3.1
Novel objects in autonomous driving
A vision system may assign high confidence when a vehicle encounters an object type absent from training, misclassifying it as a billboard or phone pole. A parallel image outlier detector can identify the object as likely out of distribution, prompting a more cautious response or a fail-safe mechanism. section 16.3
