基于改进DeepLabv3+的红外图像绝缘子与金具分割方法

Infrared Image Segmentation of Insulators and Fittings Based on Improved DeepLabv3+

摘要:

针对红外电力巡检图像中绝缘子与金具目标细长、边界模糊、局部对比度不足及背景干扰较强等问题，提出一种改进 DeepLabV3+ 的红外图像分割方法；在编码阶段引入卷积块注意力模块，增强目标区域特征表达能力，抑制复杂背景及无关热辐射信息干扰；在空洞空间金字塔池化模块中加入条带池化分支，增强模型对细长连续结构的方向性上下文建模能力，缓解绝缘子串和金具区域分割断裂问题；在解码阶段嵌入 CBAM，并结合高效通道注意力模块对融合特征进行通道重标定，进一步强化有效边界与细节信息；实验结果表明，与 U-Net、DeepLabV3+、PSPNet、SegFormer 等主流语义分割模型相比，改进方法取得了更好的分割效果，平均交并比和平均像素精度分别达到 76.13% 和 81.15%，可有效提高红外图像中绝缘子与金具目标的分割完整性和边界定位精度。

Abstract:

To address the problems of slender targets, blurred boundaries, insufficient local contrast, and strong background interference in infrared power inspection images of insulators and fittings, an improved DeepLabV3+-based infrared image segmentation method is proposed. In the encoding stage, a convolutional block attention module is introduced to enhance the feature representation of target regions and suppress interference from complex backgrounds and irrelevant thermal radiation information. A strip pooling branch is added to the atrous spatial pyramid pooling module to improve the directional contextual modeling capability of the model for slender and continuous structures, thereby alleviating the segmentation discontinuity of insulator strings and fittings. In the decoding stage, CBAM is embedded, and an efficient channel attention module is combined to recalibrate the fused features along the channel dimension, further enhancing effective boundary and detail information. Experimental results show that, compared with mainstream semantic segmentation models such as U-Net, DeepLabV3+, PSPNet, and SegFormer, the improved method achieves better segmentation performance. The mean intersection over union and mean pixel accuracy reach 76.13% and 81.15%, respectively, effectively improving the segmentation integrity and boundary localization accuracy of insulators and fittings in infrared images.