目的 近年来，利用无人机对高空输电线路巡检变得越来越普及。无人机拍摄到的图像一般具有复杂的背景，在这样的图像中进行物体的检测识别极具挑战性。方法 本文提出一种无监督的基于重复模式识别的绝缘子定位方法。在该方法中，我们使用MSER（Maximally Stable Extremal Regions）作为基本的特征，利用绝缘子伞盘之间的相似性和空间位置约束进行视觉单词的提取。在这个过程中利用了赫姆霍兹原则(Helmholtz)进行特征单词的筛选，并且使用了二值二次规划的方法进行代表特征选择。最后我们利用联合优化框架来查找具有不同的视觉物体和单词的重复模式，从而进行绝缘子的定位。我们建立了一个数据集，包含590张绝缘子图片。这个绝缘子数据集相比其他数据集具有复杂场景多、角度多，绝缘子类型多的特点。结果 相对于基准方法，基于重复模式检测的绝缘子定位方法的查全率由88.513%略微下降到82.327%的同时，准确率由19.760%上升至77.855%，表明本工作的有效性。结合直观给出的结果，能够看到本方法能够很好的应对多复杂背景下，多尺度、多角度、多类型绝缘子带来的定位困难，说明了该方法的准确性和鲁棒性。在文章中选取了SSD（Single Shot Multi-Box Detector）作为有监督方法的代表，与本文提出的无监督方法进行对比，取得了准确率85.778%，查全率88.957%的结果。一方面，验证了有监督方法在绝缘子定位上的效果；另一方面说明了本文方法作为一种无监督方法，在不引入额外数据的情况下达到了接近于有监督方法的效果。结论 本文提出的一种无监督的基于重复模式检测的绝缘子定位方法，通过实验验证了该方法在复杂背景图像中绝缘子定位的性能。
Insulator Location Based on Recurring Pattern Recognition
liyilong,zouqi(Beijing Jiaotong University)
Objective The locating and fault detection of insulators on high voltage power lines are important in transmission systems . Applying UAVs (unmanned aerial vehicles) to the inspection of high voltage power lines now becomes popular in smart transmission system. In the current UAV inspection work, the images captured by UAV are manually analyzed, which is time-consuming and manual labor-consuming. . The task of this paper is to propose a method to automatically locate insulators in aerial images. It can not only reduce manual labor cost and improve efficiency, but also benefit future fault analysis.Method Our motivation lies in the discriminative properties of insulators where each cap has a rigid form and repetitive geometric structures presented by all the caps in an insulator. Therefore, we are motivated to propose an unsupervised insulator location method based on recurring patterns. The images obtained by the UAVs are challenging in several aspects: (1) They have cluttered background, including houses, trees, rivers, towers, etc. Besides, illumination, poses and object resolutions vary greatly due to UAVs flying around the tower. Compared with images captured by a fixed camera from a specific angle, our images are more difficult to locate the insulators. (2) There are unknown number of targets (insulators) in an image and frequent occlusions, which result in considerable missing and unreliable features. Detecting and identifying the objects in such images are classic challenges in computer vision. In this paper, we propose an unsupervised insulator location method based on recurring pattern detection. The system is made up of three parts: a primitive feature extraction part, a visual word formation part and a joint optimization part. In the first part, the MSER (Maximally Stable Extremal Regions) feature is selected as the basic description, and MSER group is used for multi-scale detection. In the second part, starting from the seed selection, the primitive features extracted at the first part are grouped into visual words based on the similarity and geometric constraints. And these words are then scored by the Helmholtz principle to remove false positives, which includes the visual words in the background and the visual words not in the same directions as the true insulators. The refined words still have redundancies and are further processed by a binary quadratic programming to select representative features. In the last part, we solve a joint assignment optimization problem where different visual words and visual objects of a potential recurring pattern are considered simultaneously to locate insulators. Since there is no publicly dataset, a dataset containing 590 insulator pictures has been set up. Compared with the data set used in other literatures, this insulator data set has many characteristics, such as complex scenes, many views and resolutions and different insulator types. The experiment is carried out on the data set we established.Result In order to verify the effectiveness of our work and investigate contributions of each part to the overall system, we use three methods for comparison: 1) baselinemethod the visual words not filtered by the Helmholtz principle are directly taken as insulator location results. 2) Wordsmethod the output of visual words without the use of the joint optimization framework for recurring pattern search as insulator location. 3) Words+objectmethod take the recurring patterns after the joint optimization framework search for visual words and objects as the location of insulators. Compared with the baseline method, the recall rate of the insulators location method based on recurring patterns detection is reduced from 88.513% to 82.327%, but the accuracy rate rises from 19.760% to 77.855%, indicating the effectiveness of the work. The results show that this method can deal with the difficulties of multi - scale, multi -view angle and multi - type insulators in complex background, which shows the accuracy and robustness of the method. SSD (Single Shot Multi-Box Detector) is selected as the representative of the supervised method, compared with the unsupervised method proposed in this paper, the accuracy rate is 85.778%, and the result of the recall rate is 88.957%. On the one hand, the effect of the supervised method on the insulator location is verified. On the other hand, it is shown that the proposed method as an unsupervised method achieves the performance comparable to the supervised method without any training data. Besides, influences of different features on the system performance are investigated. Compared with MSER, SIFT and SURF features have lower signal-to-noise ratios and therefore are less suitable for insulator location in complex backgrounds.Conclusion In this paper, an unsupervised insulator location method based on recurring pattern detection is proposed. The performance of the unsupervised method in the target detection of complex background images was verified. Our collected dataset together with the annotations will be published for the related study.