Key points are not available for this paper at this time.
This paper introduces CMA-YOLO, a network that integrates multi-scale features with dual attention mechanisms to address weak feature representation, low detection accuracy, and loss of fine-grained details in deep networks for wind turbine blade surface defect detection. First, we construct the C2MSA module by designing a Multi-scale Feature-enhanced Attention Convolution Mix (MS-ACmix) based on ACmix and embedding it into the C2PSA block. This lets the network capture local and global contextual features, strengthening multi-scale target recognition and lowering missed detections. Second, we devise a Monte Carlo Dual Attention (MCDA) mechanism combining random sampling with dual attention. This approach retains the regularization benefits of the Monte Carlo method while leveraging dual attention selection, enabling improved detection accuracy with low computational cost. Finally, we substitute the original downsampling layers in the backbone and neck with the ADown module. This lightweight design, together with efficient feature extraction and fusion, reduces fine-grained detail loss and improves defect detection capability. Quantitative results reveal that, compared to YOLO11n, CMA-YOLO yields improvements of 3.4% in mAP@0.5, 6.1% in mAP@0.5:0.95, and 8.8% in recall, with a 0.7 GFLOPs reduction in computational cost, thus validating the proposed algorithm. Overall, CMA-YOLO provides a lightweight and effective approach for inspecting blade surface defects on wind turbines operating in resource-limited settings.
Li et al. (Thu,) studied this question.