Study On Image Stitching Technique Applied To The Auto Beam Inspection System Based On Machine Vision

In auto industry, the production technology for the frame beam is generally banking, punching (drilling) and molding. Because of the impact of human, die and the sate of equipments, in addition to detecting mouths and the high of wings, each batch of beams are needed to detect the number and location of important holes such as engine installation holes and steel spring stent holes. If the beam with leaks or unqualified holes appears in the producing line, severe problems will happen. Such as the frame can't be assembled or the whole truck has no methods to be assembled. Hence to detect holes is directly related to the implementation of entire assembly process. To rapidly and accurately provide detection data, it is necessary to adopt a scientific and rational method.The application of Machine vision in industrial detection becomes more active. Such as, visual inspection of printed circuit boards, and, automatic identification and classification and geometric measurement of mechanical parts. As for large components, to shoot images with a single CCD, under the premise of ensuring resolution, the vision field tends to be small. It requires to repeating several times, which limits detection rate. To increase cameras will also greatly increase system cost. Image stitching technology is to solve problems of the camera vision filed. According to sequence, it stitches high-resolution small scope images to a complete large scene image. Besides, in the process of stitching images, it amends the error information of original images caused by imaging distortion, and removes redundant information and reduces information storage, and ultimately improves detection rate and accuracy. In the paper, image stitching technology is used in machine vision inspection system, the purpose of which is to rapidly get a high-definition image of a whole beam with a single industry CCD camera, to be used to on-line detection, reducing system cost and increasing the feasibility of automatic detection of auto frame beams.In the paper, at first under the guidance of the theory of the machine vision inspection system, in terms of continuous production and detecting requirements of truck frame beam, hardware devices and software modules of the detection system is designed. The hardware acquires images and related data of the beam, and the software processes acquired images, among them image stitching module is to be studied in the paper. In order not to affect the normal production, the experiment platform is designed in the laboratory simulating scene. Detecting speed and accuracy are the two main technical indicators. They restrain each other, because the imaging area greatly detection accuracy while the number of images significantly affects detection speed. To balance the two points, the large-scale industrial CCD camera is selected in the paper.The digital image processing technology is an important theoretical basis in the paper, and image stitching technology is a part of the digital image processing technology. In the paper median filter is used to filter original images, and next aiming at images'own characteristics and detecting requirements of the beam, the algorithms in the stitching process are analyzed and improved, to optimize stitching steps:(1) The physical model is used to correct imagesSince only the beam linearly motives with the imaging system fixed including the lens has no rotation, the object distance of the adjacent images is essentially the same. The geometry deformation is mainly from the lens distortion while the distortion of the selected lens designed to large scanning sensor is smaller than that of common lenses. So the image distortion in the system can be directly corrected with simple mapping relation in the case of meeting detecting accuracy requirement. After the geometry deformation parameter is calculated, it is preserved in the data members as a parameter array. Subsequently, each image taken is geometry deformation corrected by space transform calling the parameter array.(2) The template matching algorithm based on network is used to match images.In the inspection system, because of image sequence involved in the image stitching basically only has translation on the horizontal direction without rotation, the selection of matching features doesn't need to go through complex geometric transformation. Moreover the detecting seed required in the system is relatively high. So in the paper, the template matching algorithm, which is relatively simple, is adopted to match images in the paper. Against the problems of matching speed and accuracy, in the application the following improved strategy is adopted:In the search strategy, the search region is retrained in the overlapped area. Comparing with the whole image area, it not only greatly reduces computational complexity, but also avoids error matching caused by gray similarity. In the process of testing experimental device, in terms of the beam'size and the resolution of industrial CCD camera, the work distance of the lens is determined. As the lens is fixed, the overlapped areas between image sequence are almost the same, which can be rough estimated.In the matching method, the template matching algorithm based on network is adopted. It makes use of its own characteristic which the gray difference of the adjacent two points is not large. The algorithm ensures the matching accuracy, and to a large extent reduces the computational complexity.(3) The in-and-out is used to merge images.Merging images is the last step of the course. In front, it is said one more times that the gray difference of the adjacent two images is not large. So from the perspective of computation, in-and-out method is used to merge image, and next joint seam is removed by median filter.In the end of the paper, the corresponding algorithm process is programmed by Visual C++ 6.0 on the Window XP plate and the testing result is given. In the laboratory, the stitching result and error analysis, to the steel plate whose holes are similar to the beam's, indicate that the image stitching algorithm and the program are both effective.