DLA-Net: A Dynamically Learnable Attention Network for Intelligent Surface Visual Inspection of Aero-Engine Blades

The maintenance of aero-engines is critical for ensuring the safe and reliable operation of aircraft. Given the complexity and variability of defect detection tasks in aero-engine blades, such as varying target scales, irregular spatial distributions, and environmental noise interference, this article proposes a dynamically learnable attention network (DLA-Net) to address these challenges. Specifically, in DLA-Net, a learnable branch multidomain attention (LBMA) mechanism is first developed in multiscale feature extraction and regional focusing strategy, to adaptively adjust important weights for each task-aware channel, spatial, and position. Moreover, a task-aligned dynamic detection head (TADDH) is designed, which improves the classification and localization performance by fusing task-interaction features with multiscale receptive fields. The layer-wise and multipath coordinate attention mechanisms are integrated into the TADDH, dynamically focusing on more informative layers and enhancing flexibility. In addition, a lightweight convolutional technique has been introduced to alleviate the model complexity. Finally, an enhanced focal powerful intersection over union (F-PIoU) loss is employed in the bounding box regression, where a nonmonotonic focusing mechanism is incorporated, which accelerates the convergence rate and balances the precision loss that may arise from the simplification of the model. Experimental results demonstrate the effectiveness of DLA-Net compared to other state-of-the-art (SOTA) object detection algorithms, and its compact structure indicates that the performance advantages for real-world aero-engine defect detection tasks on resource-constrained devices.