RCNet: Dual-Network Resonance Collaboration via Mutual Learning for RGB-D Road Defect Detection

Road defect detection, particularly for potholes and cracks, is a critical component of intelligent transportation systems. Deep learning methods have advanced in this field; however, existing single-network approaches face inherent challenges in addressing the differences between crack orientation sensitivity and pothole scale perception, resulting in either compromised detection accuracy or excessive architectural complexity. To address this limitation, we propose a resonant collaborative network (RCNet) framework with two lightweight specialized networks: Net1, which focuses on orientation-sensitive feature extraction in the spatial domain using a Mamba-based multidirectional perception mechanism, and Net2, which processes macro structural feature aggregation in the frequency domain using graph-wavelet transformation. To achieve effective knowledge transfer between the different networks, we introduce geometric resonance adversarial learning, which combines geometric moment constraints with conditional adversarial mechanisms to dynamically balance structural stability and discriminative capability. We further validate the generalization capability of our approach on four additional datasets. Experimental results demonstrate that the proposed RCNet outperforms state-of-the-art methods by 3.4% and 2.5% in accuracy, while requiring only 27.4% and 18.7% of the model parameters, respectively. The code is available [here].