Contextual Bias Score - A Quantitative Metric for Gender Bias Detection in Visual Datasets

Abstract

Artificial Intelligence (AI) systems in computer vision often inherit contextual biases from training datasets, resulting in unintended gender associations during visual recognition. This paper introduces the Contextual Bias Score (CBS), a novel metric for quantifying gender bias in image datasets by evaluating object prominence, 3D spatial relationships, and scene-level semantics. Unlike conventional approaches that rely on co-occurrence statistics, CBS integrates object-person spatial features, relative size, depth, and distance with CLIP-based scene embeddings to assess how visual context influences bias. The framework is evaluated across nine experimental approaches using eight COCO-derived datasets, leveraging SHAP-based interpretability, PCA-based feature weighting, and Ridge regression modeling. Approach 9 (SHAP + Ridge) achieves 85% agreement with human-annotated misclassification labels and strong score stability. Validation includes statistical significance testing, human perception alignment, and cross-approach correlation analysis. CBS offers a dataset-independent, interpretable tool for auditing contextual bias in high-impact domains. Limitations include binary gender assumptions and computational overhead (~90 seconds per image), with future work aimed at extending the framework to non-binary and culturally diverse contexts.