基于人眼视觉机制的检测机器人铸件缺陷识别与定位方法研究

发布时间：2018-06-01 19:45

本文选题：注意机制 + 射线图像　；参考：《重庆理工大学》2017年硕士论文

【摘要】：目前,X射线数字图像的计算机自动识别是射线检测领域的研究热点。从工件的X射线检测图像自动提取出其缺陷信息,传统图像识别方法均以孤立图像为研究对象进行静态图像处理和分析,自动获取被测试件的缺陷,这样导致了可靠性以及实时性不高等现实问题。机器视觉是人工智能领域一个十分活跃的分支,人眼视觉在面对复杂的环境时可以快速地将本人的注意力聚集在一些显著性高的的目标上,而后对这些目标进行优先处理。本文将人眼视觉选择注意机制引入到X射线图像目标检测中,提出一种动态的基于人眼视觉机制的X射线数字图像自动识别新方法。该方法在图像的运动序列中对缺陷进行实时提取和动态跟踪,利用人眼视觉注意机制模型等图像处理方法在每幅图像中分离出潜在缺陷,保证提取出全部真缺陷而不必考虑伪缺陷的数量,并将此方法应用于工业检测机器人在线检测过程中,有很大的现实意义。本文主要分为缺陷识别与提取、缺陷匹配与定位、系统整体设计三个部分:(1)通过模拟视觉注意机制,以区域或者行(列)分割为手段,提出了基于视觉注意机制且利用中心-周边算子的铸件射线图像缺陷检测模型以及其实现方法,避免了传统Itti方法存在的问题,减少了无关区域对核心区域的干扰以及数据处理量,达到顺利选择出射线图像中应该被关注的重点区域的目的,为后续缺陷提取打下基础。(2)分析总结了Harris角点提取和SIFT特征点提取的原理及一般流程,之后通过介绍RANSAC方法的原理及优点,优化了匹配过程中伪缺陷的问题,进一步保证了缺陷匹配的准确性。在此基础上,提出了基于Harris-SIFT的缺陷匹配方法。通过将(1)中选择的区域以及(2)中匹配成功的缺陷综合对比后,得到铸件中缺陷的真实位置,而后利用平移视差法定位铸件缺陷在空间中的具体坐标。(3)提出了基于支持向量机的铸件缺陷分类方法,构建了分类的神经网络,计算了其中的关键参数。(4)针对工业应用情况,初步设计了一套检测系统,包括系统整体的流程设计、机器人设计、相关软件的设计等。
[Abstract]:At present, the computer automatic recognition of X-ray digital image is a hot spot in the field of ray detection. The defect information is extracted automatically from the X-ray image of the workpiece. The traditional image recognition method takes the isolated image as the research object for static image processing and analysis, and automatically acquires the defect of the tested piece. This led to reliability and low real-time problems. Machine vision is a very active branch of artificial intelligence field. Human vision can quickly focus on some highly significant targets in the face of complex environment and then give priority to these goals. In this paper, the human visual selective attention mechanism is introduced into the object detection of X-ray images, and a new dynamic method for automatic recognition of X-ray digital images based on the human visual mechanism is proposed. In this method, the defects are extracted and tracked dynamically in the moving sequence of the image, and the potential defects are separated from each image by using the human visual attention mechanism model and other image processing methods. It is of great practical significance to extract all true defects without considering the number of pseudo-defects and to apply this method to the on-line detection of industrial detection robots. This paper is mainly divided into three parts: defect identification and extraction, defect matching and positioning, and the overall design of the system. The defect detection model of casting ray image based on visual attention mechanism and center-peripheral operator is proposed and its implementation method is proposed to avoid the problems of traditional Itti method. It reduces the interference of irrelevant region to the core area and the amount of data processing, and achieves the purpose of selecting the key area in the radiographic image that should be paid attention to. In this paper, the principle and general flow of Harris corner extraction and SIFT feature point extraction are summarized. Then, by introducing the principle and advantages of RANSAC method, the problem of false defect in matching process is optimized. The accuracy of defect matching is further ensured. On this basis, a defect matching method based on Harris-SIFT is proposed. The true position of the defects in the casting is obtained by synthesizing the selected areas in the casting and the matching defects in the two parts. Then using the translation parallax method to locate the specific coordinates of casting defects in space, a new method of casting defect classification based on support vector machine (SVM) is proposed, and the neural network of classification is constructed, and the key parameters. A testing system is designed, including the whole system flow design, robot design, software design and so on.
【学位授予单位】：重庆理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP242

【参考文献】