虚拟场景辅助学习的电力部件识别方法研究

发布时间：2018-05-20 18:58

本文选题：虚拟样本 + 电力部件　；参考：《武汉大学》2017年硕士论文

【摘要】：随着电力线路智能巡检的发展,采用无人机收集影像越来越多的代替了人工攀塔勘察,因此相应的关于电力设备的图像数据量也越来越大。但是由于电力方面的应用专业性强,使用范围窄,因而没有公开的相对完善标注的电力设备影像数据集。本文针对电力领域高质量标记数据不足的情况,提出了一系列利用虚拟样本结合少量人工标记样本迭代式增量学习的方法,以防振锤作为主要目标,取得了较好的结果。同时本文的方法还可以迁移应用到其他电力场景与其他电力部件中去。本文主要工作包括如下内容:1)在分析电力场景特点以及无人机拍摄工作模式的基础上,提出一种低成本的虚拟场景创建,快速生成虚拟样本及其标注的方法。2)对初始标注生成,提出一种利用基于几何约束的组合探测器(Geometrical constrait based assembled detector,GCAD)使用虚拟数据生成标注的方法,将Faster R-CNN、DPM与类Haar级联分类器综合起来。通过组合方法,结合少量的人工辅助,可显著减少数据标注的工作量。3)结合虚拟数据与真实数据的特点,开发了一个基于GCAD的人工辅助标注程序,并通过使用迭代式的Faster R-CNN训练方法,充分利用虚拟数据与真实影像中丰富的地物信息,联合训练虚拟数据与真实数据,经实验表明,可以取得较好的结果。
[Abstract]:With the development of intelligent inspection of power lines, more and more images are collected by unmanned aerial vehicles (UAVs) instead of artificial tower climbing. However, due to the strong application of power, narrow range of use, there is no relatively complete annotated power equipment image data set. Aiming at the shortage of high quality marking data in power field, a series of iterative incremental learning methods using virtual samples combined with a small number of artificial labeled samples are proposed in this paper. The main objective is to prevent vibration hammer, and good results have been obtained. At the same time, this method can also be applied to other power scenarios and other power components. The main work of this paper is as follows: (1) on the basis of analyzing the characteristics of electric power scene and the working mode of UAV shooting, this paper proposes a low-cost virtual scene creation, a method of quickly generating virtual sample and its tagging. A method of using geometric constrait based assembled detector (GCAD) based on geometric constraints to generate annotations using virtual data is proposed to synthesize Faster R-CNN DPM and Haar like cascade classifier. By combining the combination method with a small amount of manual assistance, the workload of data annotation can be significantly reduced. 3) combined with the characteristics of virtual data and real data, a manual aided annotation program based on GCAD is developed. By using iterative Faster R-CNN training method, we can make full use of the abundant ground object information in the virtual data and the real image, and jointly train the virtual data and the real data. The experiment shows that the better results can be obtained.
【学位授予单位】：武汉大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TM755

【相似文献】