OSCA-Net: An Oriented-Structure Co-Attention Network for Vision-Based Structural Damage Detection

Abstract

Visual inspection of concrete infrastructure is slow, subjective and hazardous, and the deep networks proposed to replace it are usually generic classifiers that ignore what a crack is. This paper presents OSCA-Net, a vision-based structural damage detector built around the geometry of surface distress rather than around a larger backbone. An oriented structure prior is computed without supervision from a multiscale Hessian ridge response modulated by structure-tensor coherence, so it responds to thin elongated dark features and suppresses isotropic staining. Pooled to the encoder grid, it enters every attention logit as a learnable per-head bias, guiding attention instead of cropping the image. Two frozen encoders, ResNet-50 and EfficientNet-B0, supply complementary token grids whitened to a common width by a projection fitted on training data only. Prior-biased co-attention blocks let each stream interrogate the other in both directions, and thirty-two photometric and Haralick descriptors gate the pooled visual code before classification. Only 1.04 million parameters are trained. On 40,000 concrete surface images OSCA-Net reaches 99.98 percent accuracy, 99.98 percent F1 and 0.9995 Matthews correlation, exceeding four published fine-tuned backbones on the same corpus. Clean accuracy there is saturated, every pretrained baseline falling within 1.15 points, so the models are separated instead by degradation: over four field corruptions at three severities OSCA-Net retains 98.88 percent against 97.00 percent for the weakest baseline.