SWCB-YOLO: Toward Robust and Real-Time Wind Turbine Blade Defect Detection in Unstructured Natural Environments

Technical Summary: SWCB-YOLO for Wind Turbine Blade Defect Detection

Problem Statement
Unmanned aerial vehicle (UAV) inspection is a standard strategy for wind turbine blade maintenance; however, deploying deep learning detectors in operational wind farms faces three critical challenges that degrade accuracy:

1. Unstructured Environmental Interference: Inspection images suffer from high-frequency texture interference caused by light reflections, cloud cover, and ground vegetation, leading to elevated false-positive and missed-detection rates.
2. High-Aspect-Ratio Defects: Blade cracks are often slender and bifurcating. Standard convolutional networks have limited local receptive fields, while vision transformers with square windows tend to truncate the continuous structure of these elongated targets.
3. Nonlinear Geometric Distortion: The doubly-curved, saddle-like surface of turbine blades causes nonlinear deformations when cracks are projected onto a 2D image plane. Standard bounding-box regression losses lack the geometric alignment capacity to handle these distortions, particularly in high-curvature zones like blade tips.

Methodology: SWCB-YOLO
The authors propose SWCB-YOLO, a lightweight edge detector built upon the YOLOv11n baseline, designed to operate within the computational and real-time constraints of UAV edge platforms (specifically the NVIDIA Jetson Xavier NX). The framework integrates three core components:

1. Asymmetric Strip Swin-TransformerV2 (AS-Swin2) Module:

   • Mechanism: This module replaces standard square window attention with an Asymmetric Strip Window Attention (ASWA) mechanism. It retains standard local square windows while deploying parallel vertical ($M^2 \times 1$) and horizontal ($1 \times M^2$) asymmetric partitions.
   • Function: These orthogonal strip-shaped receptive fields align with the orientation of slender cracks, capturing long-range, pixel-level dependencies along the crack axis. It incorporates a reparameterized Directional Log-space Continuous Position Bias (Log-CPB) to suppress background noise and amplify crack edge gradients.
   • Cost: It mitigates structural truncation typical of square windows with minimal added computational cost.

2. Frequency-Spatial Dual-Domain Attention (FS-DDA):

   • Mechanism: A low-overhead module that cascades 2D Discrete Cosine Transform (2D-DCT) spectral channel weighting directly into a large-receptive-field spatial attention stage on a single backbone.
   • Function: It decouples high-frequency environmental clutter (e.g., vegetation, reflections) from weak defect edges. The spectral branch attenuates high-frequency noise, while the subsequent spatial branch refines geometric continuity. This dual-domain approach separates low-frequency backgrounds from high-frequency abnormal edges more effectively than channel-only attention mechanisms.

3. Curvature-Aware Shape IoU (CA-Shape-IoU) Loss:

   • Mechanism: A training-only localization loss that derives supervision from the underlying crack morphology rather than just the axis-aligned bounding box.
   • Components: It utilizes precomputed offline fields from ground-truth polygon masks:
     • Bézier-curvature weights: Apply higher gradient penalties in high-curvature regions (e.g., blade tips) to correct drift.
     • Voronoi skeleton-distance field: Pulls predicted box boundaries toward the true crack skeleton to reduce false positives at bifurcations.
     • Elongation constraint: Enforces scale accuracy for targets with extreme aspect ratios ($r > 8$).
   • Inference Impact: Because the morphological terms are precomputed and applied only during training, the detection head remains unchanged, and there is zero additional inference cost.

Key Contributions

1. Morphology-Derived Training Loss: The paper introduces CA-Shape-IoU, a geometric alignment criterion that supervises axis-aligned detection boxes using precomputed curvature and skeleton-distance fields. This is distinct from prior shape-aware losses that rely on box geometry alone or segmentation-based skeleton losses that require inference-time segmentation heads.
2. Task-Specific Anisotropic Backbone: The AS-Swin2 module specifically addresses the truncation of slender defects by combining standard square windows with orthogonal strip partitions and direction-adaptive positional encoding within a SwinV2 block.
3. Low-Overhead Dual-Domain Refinement: The FS-DDA mechanism achieves frequency-spatial separation without a second frequency backbone, cascading spectral weighting into spatial attention to maintain a single-backbone, low-latency design.
4. Rigorous Evaluation Protocol: To ensure fair comparison, all baseline and competing detectors were re-trained from scratch under an identical protocol (same data splits, input resolution, optimizer, and hyperparameters) on a field dataset spanning a full annual cycle. The method was also validated on three open benchmarks (NEU-DET, DAGM, VisDrone-2019).

Results
Evaluated on a proprietary field dataset of 5,707 images from a 1.5MW wind farm:

• Accuracy: SWCB-YOLO achieved 89.0% mAP@50 and 66.4% mAP@50:95. This represents an improvement of 13.8 percentage points in mAP@50 and 25.3 percentage points in mAP@50:95 over the YOLOv11n baseline.
• Efficiency: The model runs at 35 FPS on an NVIDIA Jetson Xavier NX with 3.70M parameters and 8.70 GFLOPs.
• Ablation Insights:
  • The AS-Swin2 backbone contributed the majority of the gain (~87% of the mAP@50 improvement).
  • FS-DDA provided complementary refinement, particularly in vegetation-heavy and hazy conditions, recovering recall lost by frequency filtering alone.
  • CA-Shape-IoU specifically improved high-IoU localization (mAP@50:95), correcting geometric drift without increasing latency.
• Cross-Dataset Generalization: The model demonstrated positive transfer to NEU-DET (+7.2% mAP), DAGM (+4.6% mAP), and VisDrone-2019 (+3.8% mAP), though gains decayed as domain distance increased, consistent with the specialized nature of the modules.
• Statistical Significance: Paired t-tests and Wilcoxon signed-rank tests confirmed that the performance improvements over baselines are statistically significant ($p < 0.001$).

Significance and Claims
The paper claims that SWCB-YOLO addresses the specific gap of detecting minor, elongated defects on doubly-curved surfaces in unstructured natural environments while maintaining real-time inference on edge devices.

• Robustness: By integrating frequency-domain filtering and anisotropic attention, the model effectively suppresses false positives caused by complex backgrounds (vegetation, reflections) that typically plague standard detectors.
• Geometric Precision: The CA-Shape-IoU loss demonstrates that incorporating offline morphological priors (curvature and skeleton distance) can significantly improve bounding box alignment for curved targets without incurring inference penalties.
• Practical Deployment: The framework balances accuracy and resource constraints, offering a viable solution for automated UAV-based blade inspection where high precision is required alongside strict latency and power budgets.

The authors note limitations, including the dependence on polygon-level annotations for the CA-Shape-IoU loss (though the method remains robust to moderate annotation simplification) and potential over-smoothing in scenarios combining extreme motion blur with fog. Future work is directed toward reducing annotation dependency via semi-supervised learning and exploring multimodal fusion.