GMCF-lite for lightweight and interpretable traffic sign classification via gated mamba-CNN fusion

This paper presents GMCF-Lite, a novel lightweight architecture that integrates the long-range modeling capabilities of State Space Models (Mamba) with the local feature extraction strengths of Convolutional Neural Networks (CNN) through two core innovations: the Bidirectional Interaction Module (BIM) and Adaptive Gated Fusion (AGF). Unlike conventional fusion strategies that rely on simple concatenation or element-wise addition, the proposed approach enables dynamic mutual modulation between parallel streams, allowing each branch to selectively enhance or suppress features based on the other's contextual state. To ensure reproducibility, we report results averaged over 5 independent training runs with different random seeds, and provide detailed architecture configurations and training protocols. Extensive experiments on the German Traffic Sign Recognition Benchmark (GTSRB) and a filtered subset of the Tsinghua-Tencent 100K (TT100K) datasets demonstrate that GMCF-Lite achieves 99.5% Top-1 accuracy on GTSRB and 94.5% on TT100K, outperforming state-of-the-art lightweight models including MobileViT, EdgeViT, ConvNeXt, MobileNetV3, and RepVGG by 1.8-2.7% while maintaining only 8.7M parameters and 0.45G FLOPs. Comprehensive ablation studies validate the individual contributions of BIM and AGF, showing performance drops of 1.8% and 2.1% respectively when removed. Additional robustness evaluations under blur, occlusion, and low-light conditions confirm the architecture's practical reliability. Visual interpretability analysis through Grad-CAM with quantitative IoU metrics reveals that GMCF-Lite produces sharply focused attention maps concentrated on discriminative sign regions, in contrast to the noisy local activations of CNNs and diffuse background responses of pure Mamba variants.