Abstract: High-voltage cables are widely used in power systems and are important carriers of power transmission,and the detection of the number of cable cores is of great significance for the safe operation of power systems.The common core count detection method is based on the cable cross-section,and the cable core cross-section texture has a great influence on the detection results.To address this problem,a wire core cross-section texture elimination method based on the improved Pix2Pix (Pixel-to-Pixel) model was proposed,which could effectively remove the texture of the wire core cross-section.U-Net was used as the generator and PatchGAN was used as the discriminator to complete the transformation of the wire core cross-section image.To improve the accuracy of the image generation,the self-attention mechanism and group convolution were added in the up-sampling stage of the model,so that the lightweight model could ensure the integrity of the wire core conductor contour while eliminating the texture.Finally,a watershed segmentation algorithm was used to accurately and efficiently segment and calculate the number of cable core roots.The method was tested in the cable datasets,and the experimental results showed that the texture elimination rate reached 99.2%,and the correct detection rate of the number of wire cores reached 98%.Compared with the traditional de-texturing methods and mainstream generative adversarial networks such as U-GAN and StyleGAN,the Pix2Pix-SA network proposed was more effective in eliminating the texture of high-voltage cable wire core cross-section,and improved the efficiency of cable wire core number detection.

Key words: high voltage cable, Pix2Pix, self-attention mechanism, mark texture elimination, image generation

摘要： 鉴于高压电缆线芯数量的准确检测对电力系统安全的重要性,而截面纹理显著影响检测结果,提出一种基于改进Pix2Pix(Pixel-to-Pixel)模型的线芯截面纹理消除方法,可有效去除线芯截面的纹理。运用U-Net作为生成器,PatchGAN作为判别器完成对线芯截面图像的转换。为了提高图像生成时的精度,在模型上采样阶段加入自注意力机制和分组卷积,使轻量化模型在消除纹理的同时保证线芯导体轮廓的完整性。最后采用分水岭分割算法准确、高效地分割计算电缆线芯根数。通过在电缆数据集中进行测试表明,所提方法的纹理消除率达到99.2%,线芯数量正确检测率达到98%。相比传统去纹理方法和U-GAN,StyleGAN等主流生成对抗网络,所提Pix2Pix-SA网络对高压电缆线芯截面纹理的消除效果更好,同时提升了电缆线芯数量的检测效率。

关键词: 高压电缆, Pix2Pix, 自注意力机制, 纹理消除, 图像分割