Fine-grained recognition method for corrosion areas on surface of transmission equipment based on image semantic segmentation

CHEN Bojian; WU Wenbin; LIN Chenghua; LIANG Manshu; WU Xiaojie

doi:10.7688/j.issn.1000-1646.2025.03.10

Abstract：[Objective]With the continuous expansion of power grid scale and the increasing complexity of the operating environment,surface corrosion of transmission equipment has become a critical factor threatening the safe operation of power grids.Traditional manual inspection methods are not only inefficient but also struggle to accurately identify subtle corrosion features on equipment surfaces,especially in complex natural environments where the boundaries of corrosion areas are often blurred,posing significant challenges for precise recognition.To address this,a fine-grained recognition method for corrosion areas on the surface of transmission equipment based on image semantic segmentation was proposed,aiming to achieve precise detection and recognition of corrosion areas through deep learning technology.[Methods]The core of this method was the construction of a semantic segmentation network integrated with an attention mechanism.This network,by introducing both channel attention and spatial attention mechanisms,could effectively capture the subtle features and precise boundaries of corrosion areas.Specifically,the channel attention mechanism enhanced the response to channels with prominent corrosion features by analyzing the relationships among various channels in the feature map.Meanwhile,the spatial attention mechanism strengthened the spatial feature representation of corrosion areas by focusing on the spatial location information in the feature map.After the initial segmentation,the K-means++clustering algorithm was employed to perform clustering analysis on the RGB values of the pixels in the segmented images.By optimizing the selection of initial clustering centers,this algorithm effectively avoided the issue of local optimum that could arise with the traditional K-means algorithm,thereby more accurately dividing corroded and uncorroded areas.To further improve recognition accuracy,the structural similarity index was introduced to evaluate each clustered area,and fine-grained recognition of corrosion areas was achieved at the pixel level by calculating the structural similarity between areas.[Results]Experimental results demonstrate that the proposed method exhibits remarkable performance on a dataset of transmission equipment images in complex natural environments,achieving a significantly improved corrosion area recognition accuracy and an obvious improvement in boundary localization accuracy compared to traditional methods.[Conclusion]In summary,the semantic segmentation network integrated with an attention mechanism,combined with the K-means++clustering algorithm and SSIM evaluation,pioneers an efficient and precise new approach for fine-grained recognition of corrosion areas on the surface of transmission equipment.By incorporating the attention mechanism,the proposed method effectively addresses the challenges posed by complex corrosion features and blurred boundaries,significantly enhancing recognition accuracy.Meanwhile,the combination of the clustering algorithm and SSIM evaluation enables pixel-level detailed differentiation,further improving the fineness and practicality of recognition and providing solid technical support for the safe monitoring and maintenance of power grids.Not only does the proposed method ensure the safe and stable operation of power grids,but it also offers valuable insights and inspiration for the application of image recognition and segmentation technologies in other fields.