Surface damage feature segmentation and extraction of wind turbine blades based on superpixel dynamic threshold

Machine vision-based monitoring is increasingly employed to detect surface damage on wind turbine blades (WTB). However, challenges such as complex natural backgrounds and non-uniform illumination often compromise detection accuracy. In this work, we propose a superpixel-based dynamic threshold segmentation method to enhance the accuracy and robustness of blade damage detection under complex conditions. Firstly, we introduce a Multi-Scale Retinex feature enhancement algorithm with histogram constraints (HC-MSR). By incorporating histogram clipping constraints into the MSR algorithm, our approach effectively corrects non-uniform illumination and enhances the contrast of damage features on blade surfaces, thereby preserving essential image details. Subsequently, we utilise the SLIC algorithm to generate superpixel semantic units, enabling decoupling of heterogeneous features. A dynamic threshold segmentation model based on local luminance distribution is proposed to segment the damage features while minimising residual natural background. Finally, any remaining background components are eliminated through background mask replacement. Experimental results demonstrate that the proposed method significantly outperforms existing techniques in detecting surface damage on WTB under complex operating conditions. Compared to the average performance of K-means+Iterative threshold, K-means+OTSU, and FCM methods, the proposed approach improves accuracy, IoU, and F-measure by 41.99%, 52.54%, and 61.99%, respectively, demonstrating greater robustness and practical applicability.