MSFBHNet: A multiscale spatial-frequency Bayesian-inspired hybrid network for concrete surface defect identification

Abstract Concrete surface defect identification is crucial for infrastructure safety assessment and subsequent maintenance. However, existing methods still exhibit instability when faced with challenges such as weakened texture and strong noise interference. To address this, we propose MSFBHNet, a novel multiscale spatial-frequency Bayesian-inspired hybrid network designed for robust identification of concrete surface defects. Specifically, we first extract morphological variation information under different receptive fields through the Multiscale Asymmetric Convolution (MAConv) block, enabling the model to effectively represent concrete surface damage features at different scales and morphologies. Subsequently, we design a Global-Directional Frequency Attention (GDFA) module to capture global spectral semantics and directional energy distribution features, and utilize this information to adaptively modulate spatial features, thereby enhancing the model's robustness under complex backgrounds and noise interference. Furthermore, we introduce a Bayesian-inspired Distributional Prediction (BDP) module featuring learnable Gaussian classifier parameters. Reparameterization sampling is employed during training, while Monte Carlo inference aggregates predictions from multiple sampled classifiers to obtain more robust predictions. Extensive experiments on three authoritative public datasets indicate that the proposed method achieves superior results in multiple metrics, including recall, F1 score, and accuracy, while requiring only 3.88 million parameters and 0.46 GMACs. This demonstrates that the proposed method achieves a favorable balance between identification performance and computational efficiency.