捆扎线材图像处理的识别方法研究

【作者】 陈浩；

【导师】 王景中；

【作者基本信息】 北方工业大学 ， 计算机应用技术， 2004， 硕士

【摘要】 目前，国内大多数钢材生产厂家都采用人工方法来计量捆扎线材的根数，效率低下，劳动强度高，计数误差大，而引进全自动生产线，成本较高，因此，他们迫切需要低廉的自动计数仪。从上世纪90年代初，国内开始从事应用计算机视觉的方法进行线材的在线自动计数研究，取得了一定的成果，但还处于理论研究阶段，没有出现成熟的产品。本课题是北方工业大学承担的北京市教育委员会科学基金的捆扎线材图像测量计数仪的子课题。该计数仪的研制不仅具有一定的理论意义，同时还有良好的市场前景。 课题仍然采用计算机视觉的方法来尝试线材的在线计数，主要研究内容是系统的软件部分，包括图像的采集、预处理、物体的分割和识别并且计数。根据所处理图像的实际特点，文中提出了一种新的基于类圆的粘连体分割方法和基于统计模式识别的类圆识别方法，由于该算法主要的运算为加法、减法和逻辑运算，从而保证了算法的实时性。对一幅大小为640~*480的图像，整个处理耗时小于1秒，完全能满足实际生产要求。该算法对图像的二值化效果要求较高，当物体间的缝隙在二值化后能部分或全部判定为背景，即粘连的物体在二值化后能有明显的凹陷时，识别率很高，误差几乎可以为零。反之，则误差会上升，一种可行的改进方案是利用物体的边缘信息，文中也进行了这方面的尝试，受时间限制，只进行了简单的试验。虽然引入边缘信息能减少可能的误差，但是如果物体内部灰度不均匀，则要考虑由此而造成的伪边缘信息对分割的影响，同时，结合边缘信息后，处理速度也将会有所降低。 文中的研究大部分时间是针对用数码相机在施工现场采集的图像设计的，仅在末期用了较少的几根钢筋进行在线测试，因此，仅就算法而言，还有许多具体的研究、工作去调试、完善和改进，如果再进一步结合嵌入式硬件和生产管理的需要，形成一个较成型的产品，则需要做更多、更深入的研究。 文中除对算法进行了描述外，也给出了算法的应用效果，并且简要地讨论了引起误差的几个因素。 本文在查阅大量的文献基础上(限于作者查找的资料范围，没有见到国外有关这方面的文献)，提出了一种适合线材计数的分割和识别方法，并就进一步的改进进行了一些有益的探索，为日后的计数仪研制打下了坚实的基础。更多还原

【Abstract】 At present, almost all steel factories count bundled bar one by one by workers, it is laborious, low effective, and may cause high error. If these factories would import foreign automatic product lines, they will pay much and improve the cost of steel bar dramatically. So, these manufactories desire a cheap, applied counter. From the first of 1990s, by using computer vision, studies of this area had been begun. Although a bit progress had been got since then, there is much to do if it is used at spot in factory. This study is sub study of counter of bundle bar by image processing, taken on by North China University of Technology, sponsored by education committee of Peking city. This study may do some research on theory and the product will be a good future.In this paper, the author tries to count steel bars of each bundle by making use of computer vision. The main research includes the following area: image acquisition, image pre-processing, image segmentation, object recognition. An new segmentation method based on quasi-circular assumption and an new object recognition method based on scanning are present. Owing to the simple addition, subtraction and logic operation, these two methods have real time property. The total CPU time is less than 1 second for an 640*480 pixels image, this can meet on line counting. Both two methods require a good binary image, if there exist concave, the aggregated objects will be segmented and recognized correctly and the error is lower, otherwise, it may give err result. Considering the edge information will give robust segmentation, but the information may contain noise when the object is strongly non-uniformity and the speed decreases.The author pay much time to test the arithmetic on the image photographed by digital camera, so there are much works to do to apply this software in practice.In this paper, not only do we explain our new algorithm, but also we test them by real digital photo, man-made image, video image. Furthermore, we compute and discuss the error and its factors. After consulting many references, this paper presents a fast method to count the steel bar of one bundle.更多还原