基于语义分割数据增强与可变形卷积的输变电缺陷检测

Defect Detection for Transmission and Transformation Lines Based on SAM Data Augmentation and Deformable Convolution

摘要:

为提高输变电线路在长期运行和自然环境影响下产生的各类缺陷的检测精度和效率，提出了基于YOLOv8和可变形卷积的改进算法。针对各类缺陷样本数量差距较大导致的长尾分布问题，提出使用分割一切模型对数据进行扩充以提高数据平衡性；在主干网络中引入DCNv2结构，通过动态调整卷积核的形状适应各类样本包含的特征以提高泛化能力，并结合multiCA注意力机制使网络注重各通道的融合信息；使用损失函数WIoU引导模型学习，以适应不同质量的锚框。将改进的算法与其他算法进行比较，结果表明改进算法增加了对输变电缺陷检测的精度。

Abstract:

In order to enhance the detection precision and efficiency of various defects in transmission and transformation lines under the influence of long term operation and natural environments, an improved algorithm based on YOLOv8 and deformable convolution is proposed. Aiming at long tail distribution problem caused by the significant disparity in the quantity of various defect samples, the segment anything model (SAM)is proposed to augment the data so as to enhance data balance. Within the Backbone, DCNv2 is introduced to dynamically adjust the shape of convolutional kernels to adapt to the features contained in various samples, thus enhancing generalization capability. Moreover, the integration of multi CA attention mechanism guides the network to focus on the fusion information of each channel, and WIoU is utilized to guide the model learning, which enables adaptation to anchor boxes of different qualities. Comparative analysis with other algorithms shows an increase in detection accuracy for the defects in transmissi on and transformation lines achieved by the proposed improved algorithm.