Fuzzy-DogNet: A Lightweight and Interpretable Deep Unstructured Neuro-Fuzzy System Based on Band-Pass Filters for Traffic Sign Recognition

Abstract

Traffic Sign Recognition (TSR) plays a critical role in autonomous driving and advanced driver-assistance systems, necessitating accurate and efficient recognition under real-world constraints. This paper introduces Fuzzy-DoGNet, a novel lightweight architecture designed to enhance TSR performance while addressing challenges such as environmental variability, sensor noise, and resource limitations. The proposed model integrates a learnable Difference-of-Gaussian (DoG) module for adaptive, edge-enhanced feature extraction, a multi-scale feature pyramid to achieve scale invariance, and a feature-map attention mechanism for effective feature aggregation. These components feed into an Unstructured Neuro-Fuzzy Inference System (UNFIS), enabling robust decision-making under uncertainty. By emphasizing structural elements such as edges and line-based features, Fuzzy-DoGNet enhances recognition robustness in challenging conditions. Experimental evaluations demonstrate that the proposed architecture achieves competitive accuracy with only 1.9 million parameters, outperforming several state-of-the-art models in terms of efficiency and robustness, thereby making it suitable for real-time deployment on resource-constrained platforms.