毕业生知识库

绝缘子在高压输电线路中起着至关重要的作用。绝缘子长期运行在野外，极易受外界环境影响而损坏，影响电力的正常输送，因此需要定期检查。利用无人机等携带照相设备进行检查，是目前常用的方法。而通过上述方式获取航拍图像后对绝缘子等部件的故障检测仍采用人工方式进行，不但耗时耗力且检测准确率易受到人为因素影响。利用视觉技术进行绝缘子故障的自主检测是提高检测效率和准确率的重要途径，而航拍图像中绝缘子的自动定位是实现绝缘子故障自主检测的前提，因此具有重要研究意义。

本文以高压输电线路自主巡检为研究背景，以实现复杂环境下高压输电线路航拍图像中多姿态绝缘子的自动检测为目标。通过基于SVM的绝缘子检测方法和基于语义分割网络的绝缘子目标分割方法，实现了航拍图像中绝缘子的自动检测。论文的主要工作如下：

1、针对利用航拍图像进行输电线路故障检测问题，本文对基于图像的绝缘子检测方法进行了综述和分析，指出了现有方法存在的复杂背景下检测难度大、多姿态绝缘子检测结果不精确、检测准确率低等问题；对目前常用的基于卷积神经网络的目标检测方法进行了综述和分析，指出了其在解决此类问题上的优势；分析了航拍图像中绝缘子检测的关键技术问题并介绍了论文的结构安排。

2、针对复杂背景下多姿态绝缘子的精确检测问题，提出了一种基于SVM的绝缘子检测方法。该方法采用HOG和LBP的融合特征，提高了检测准确率；通过图像预处理等方法，减少候选子窗口数，提高了检测速度；采用非极大值抑制和子窗口合并方法，得到多姿态绝缘子所在倾斜矩形框的位置，解决原始SVM检测方法绝缘子检测结果不精确的问题，最终实现多姿态绝缘子的快速准确检测。

3、为进一步提高绝缘子检测速度和精度，提出了一种基于语义分割的绝缘子检测方法。该方法对FCN和U-Net网络进行对比，并对U-Net网络进行改进。在U-Net网络中，使用不对称的小卷积核代替大卷积核，减少参数量，提高检测速度；采用数据增强和参数正则化方法，提高了网络泛化能力；采用通道裁剪和量化的方法，减少了参数量和模型尺寸，提高了检测速度。最后对通过分割网络得到的预测图像进行处理最终实现绝缘子检测。

Insulators play a vital role in high voltage transmission lines. Insulators work in the field for a long period of time and are easily damaged by external environmental influences, affecting normal power transmission. Therefore, regular inspections to insulators are required. Using drones by carring portable camera devices for inspection is a common method at present. However, the fault detection of insulators and other components after obtaining the aerial image through the above method is still performed manually, which is not only time consuming and labor-consuming but also the detection accuracy rate is easily affected by human factors. The use of visual technology for auto-detection of insulator faults is an important way to improve the detection efficiency and accuracy. The auto-positioning of insulators in aerial images is the premise to achieve auto-inspection of insulator faults, so it has important research significance.

This paper takes the self-inspection of high-voltage transmission lines as the research background, and aims to achieve the automatic detection of multi-pose insulators in aerial images of high-voltage transmission lines in complex environments. Insulator detection method based on SVM and insulator segmentation based on semantic segmentation network are used to achieve automatic detection of insulators in aerial images. The main work of the dissertation is as follows:

1. Aiming at the problem of transmission line fault detection using aerial images, this paper summarizes and analyzes the image-based insulator detection methods, and points out that the existing methods have problems such as difficult to detection under complex backgrounds, inaccurate detection results of multi-pose insulators, and low detection accuracy; reviewed and analyzed the commonly used target detection methods based on convolutional neural networks, pointed out its advantages in solving such problems; analyzed the key technical issues of insulator detection in aerial images, and introduced the structure of the thesis.

2. Aiming at the problem of accurate detection of multi-pose insulators under complex background, an insulator detection method based on SVM is proposed. This method adopts the fusion feature of HOG and LBP to improve the detection accuracy; image preprocessing method is used to reduce the number of candidate sub-windows and improve the detection speed; The non-maximum suppression and sub-window merge methods are used to obtain the position of oblique rectangle of the multi-pose insulators, solving the problem of inaccurate insulator detection results in the original SVM detection method, and finally achieving the rapid and accurate detection of multi-pose insulators.

3. To further improve the detection speed and accuracy, an insulator detection method based on semantic segmentation is proposed. This method compares FCN and U-Net networks and improves the U-Net network. In the U-Net network, asymmetric small convolution kernels are used instead of large convolution kernels to reduce the number of parameters and improve detection speed; Data enhancement and parameter regularization methods are used to improve network generalization capabilities; channel clipping and quantization are used, which can reduce the amount of parameters and the model size and improve the detection speed. Finally, we process the predicted image obtained by segmentation network to finally achieve the insulator detection.

4. For the two methods proposed in this paper, they were respectively verified by experiments, the results show that both methods can achieve the detection of multi-pose insulators. Finally, compared the two methods, and experimental results show that detection method based on semantic segmentation has better detection accuracy and detection speed.