Asymmetric similarity aware object detection using Tversky YOLO for fine grained traffic sign recognition

Abstract

Abstract Traffic sign detection requires fine-grained discrimination among visually similar categories under occlusion, illumination variation, and extreme scale changes. Most modern detectors rely on symmetric inner-product classifiers that penalize missing expected features and extraneous activations equally. This assumption is poorly aligned with traffic sign semantics, where the absence of a diagnostic symbol is often more informative than incidental background responses. We propose Tversky-YOLO , a YOLO11m-based detector that embeds asymmetric similarity directly into the detection pipeline. TverskyFusion modulates backbone residual connections using a local Tversky similarity map to suppress context-inconsistent activations, while TverskyProjection replaces the terminal 1×1 classifier with a prototype-based asymmetric similarity operator whose log-odds outputs remain compatible with binary cross-entropy training. Together, these components reshape classifier decision geometry by inducing directionally weighted margins while keeping the parameter count nearly unchanged relative to the YOLO11m baseline. To isolate the contribution of asymmetry from prototype capacity and generic gating, we introduce symmetric and sigmoid-based control conditions and evaluate all models across five independent seeds with paired statistical testing. On GTSDB, Tversky-YOLO achieves 96.8 ± 0.4 and 82.1 ± 0.6% mAP@ 0.5:0.95, significantly surpassing a YOLO11m baseline. On TT100K, it attains 96.3 ± 0.5% and 74.4 ± 0.7%, respectively, with gains broadly distributed across categories.