电路板残片图像识别

发布时间：2018-03-18 05:38

本文选题：对象分割　切入点：概率图模型　出处：《北京邮电大学》2017年硕士论文　论文类型：学位论文

【摘要】：爆炸犯罪已成为近年来公共安全面临的最大威胁之一,其中尤以遥控/定时等非现场智能启动的方式危害最大。研发此类装置中核心部件-电子线路板炸后识别比对方法并进一步溯源追踪,为侦破此类爆炸案件提供重要物证线索。同时,近年来智能视频监控在公共安全领域发挥着不可替代的作用,智能视频分析中的关键技术包括了目标跟踪与事件检测。本文主要对电路板图像分割、图像识别以及监控视频群体事件检测进行了深入研究,取得的主要成果如下:图像分割方面,本文提出并实现了一种基于概率图模型与概率抽样颜色模型的电路板对象分割方法。该方法基于电路板自身的特点,采用分层分割的策略对电路板图像进行元器件分割。在电路板整体分割层面,利用K-means聚类找出主板候选像素,并分别初始化前景(元器件对象)与背景(主板)的高斯混合模型,进而使用Grab cut算法反复迭代分割;在元器件局部分割层面,基于Grab cut分割结果,通过均匀抽样得到前景(元器件自身)与背景(主板及周围元器件)的颜色模型,使用最近邻思想对像素分类,精细修正元器件分割结果。该方法在公开的电路板数据集以及我们自行采集的数据集中均取得了良好的分割效果,对图像的阴影、噪声具有鲁棒性。图像识别方面,本文实现并改进了一种基于SIFT特征几何相似性的特征点匹配方法,用以解决电路板残片图像识别问题。该方法利用三角形结构稳固的特性,在匹配特征点描述时以空间相似三角形加以约束,有效地减少了特征点错误匹配。针对电路板图像识别任务,本文对该方法进行了关键性改进,修改特征点对的筛选策略,同时增加了更强的约束,进一步提升了识别准确率,改进方法在实验测试中准确率达到99.1%。事件检测方面,本文改进了一种多目标跟踪方法,并设计实现了一种新的轨迹分析方法检测群体事件。通过引入高斯过程回归,拟合底层连接的轨迹片段以及高层连接中非线性运动模式,提高了跟踪的准确率。基于跟踪的轨迹,提取运动速度、运动方向、轨迹间距离等特征,通过分析轨迹间的关系,判断是否有分离或者聚集事件发生。在2015年的TRECVID-SED评测中,该方法用于人群分离(PeopleSplitUp)和聚集(PeopleMeet)事件检测,并取得了排名第2和第4的成绩。
[Abstract]:The crime of bombing has become one of the greatest threats to public safety in recent years. In particular, off-site intelligent startup, such as remote control / timing, is most harmful. At the same time, intelligent video surveillance has played an irreplaceable role in the field of public security in recent years. The key technologies of intelligent video analysis include target tracking and event detection. The main achievements are as follows: in the aspect of image segmentation, this paper proposes and implements a method of PCB object segmentation based on probabilistic graph model and probabilistic sampling color model, which is based on the characteristics of the circuit board itself. In the whole segmentation level of circuit board, K-means clustering is used to find the candidate pixels of the main board. The Gao Si hybrid model of foreground (component object) and background (motherboard) is initialized, and then repeated iterative segmentation using Grab cut algorithm. At the local level of component segmentation, the segmentation result is based on Grab cut. The color models of foreground (components themselves) and background (mainboard and surrounding components) are obtained by uniform sampling. The nearest neighbor idea is used to classify pixels. The method has good segmentation effect on the open circuit board data set and the data set we collected by ourselves, and is robust to the shadow and noise of the image. In this paper, a feature point matching method based on geometric similarity of SIFT features is implemented and improved to solve the problem of image recognition of circuit board fragments. The matching feature points are constrained by spatial similar triangles, which can effectively reduce the error matching of feature points. In view of the task of image recognition of circuit boards, this paper makes a key improvement on the method and modifies the selection strategy of feature points. At the same time, it adds stronger constraints, further improves the recognition accuracy, the improved method in the experimental test accuracy reaches 99.1.The event detection, this paper improved a multi-target tracking method, A new trajectory analysis method is designed and implemented to detect group events. By introducing Gao Si process regression, the trajectory segment of the bottom link and the nonlinear motion pattern in the high-level connection are fitted. The tracking accuracy is improved. Based on the track tracking, the characteristics of velocity, direction and distance between tracks are extracted, and the relationship between tracks is analyzed to determine whether there are separation or aggregation events. In the TRECVID-SED evaluation in 2015, The method was applied to the detection of people split up (PeopleSplitUpp) and crowd Meet (PeopleMeet) events, and the results were ranked 2nd and 4th.
【学位授予单位】：北京邮电大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：D917.9;TP391.41

【相似文献】