SceneGraphSafetyClassifier<T>

Models & Types · AiDotNet.Safety.Image

Scene graph-based image safety classifier that analyzes spatial relationships between detected entities to identify unsafe content configurations.

Inheritance: ModelBase<T, Vector<T>, Vector<T>> → SafetyModuleBase<T> → ImageSafetyModuleBase<T> → SceneGraphSafetyClassifier<T>

For Beginners

This classifier works like a detective examining a scene: first it identifies what objects/regions are present (skin, dark areas, bright areas), then checks how they relate to each other spatially. For example, a weapon-shaped object near a person-shaped region is more concerning than either alone.

How It Works

Rather than classifying the image as a whole, this classifier segments the image into regions, characterizes each region’s visual properties, and then analyzes spatial relationships between regions. Certain spatial configurations (e.g., skin-colored regions in specific arrangements, weapon-shaped objects near person-shaped regions) indicate unsafe content that whole-image classifiers may miss.

References:

  • USD: Scene-graph-based NSFW detection for text-to-image (USENIX Security 2025)
  • Scene Graph Generation survey: methods, challenges, applications (2024)
  • OmniSafeBench-MM: 9 risk domains with 50 fine-grained categories (2025, arxiv:2512.06589)

Constructors

Constructor Summary
SceneGraphSafetyClassifier(double threshold, int gridSize) Initializes a new scene graph safety classifier.

Properties

Property Summary
string ModuleName { get; }

Methods

Method Summary
IReadOnlyList<SafetyFinding> EvaluateImage(Tensor<T> image)