TDDC-YOLO: texture–defect disentanglement for robust wood surface crack detection under domain shifts

Wood surface crack detection is crucial for automated quality inspection in wood processing, yet remains challenging in real production lines due to thin and low-contrast crack patterns, strong interference from repetitive wood-grain textures, and pronounced appearance shifts across wood species and illumination conditions. To address these issues, we propose TDDC-YOLO, a robust crack detection framework built upon YOLOv8. The core idea is to explicitly disentangle texture and defect representations via an orthogonality-constrained decoupling component, thereby suppressing texture-induced false activations. Meanwhile, we introduce a lightweight crack-geometry guidance branch with readily constructible pseudo labels to enhance the structural consistency of slender cracks, and further improve robustness under domain shifts using a frequency-aware mixing strategy with matched-positive consistency regularization. Experiments on the evaluated wood-surface crack dataset show that TDDC-YOLO improves detection accuracy with limited additional overhead, reaching [Formula: see text] = 0.938 and [Formula: see text] = 0.673 on the test set while maintaining real-time inference on our test platform (3.9 ms/image, 256 FPS). Moreover, cross-wood evaluations on two species (Ash and Bubinga) indicate improved robustness to wood-species induced texture shifts (e.g., Ash-to-Bubinga [Formula: see text] from 0.575 to 0.615), suggesting its potential for more stable detection under the evaluated cross-wood setting.