YOLO with Kolmogorov-Arnold networks and vision-language foundation models for interpretable object detection with trustworthy multimodal AI in computer vision perception

The researchers proposed a new computer vision framework that applies the Kolmogorov-Arnold (KAN) network to the You Only Look Once (Yolov10) model to enhance the reliability of object detection. This framework uses KAN as an interpretable posterior proxy model, combining seven geometric and semantic features to model detection results. It achieves direct visualization of feature effects through additive spline structures, generating smooth and transparent functional mappings to reveal the reliability of confidence estimates. Additionally, the study introduced the Bootstraking Language Image (BLIP) base model to generate descriptive titles for scenarios, constructing a lightweight multi-modal interface without interfering with the interpretability layer. Experiments on the COCO general object dataset and Bask University campus images demonstrated that this framework can accurately identify low-confidence predictions under blurred, occluded, or low-textured conditions, providing actionable insights for manual review and downstream risk mitigation. Ultimately, it achieved an interpretable object detection system with credible confidence estimates.