Abstract: A machine vision system for bearing outer diameter inspection was developed to meet the requirement of online inspection. Higher inspection precision was obtained due to the introduction of simple and precise calibration method along with effective and accurate image-processing algorithm (ex . Spatial moment sub-pixel edge location algorithm) . The experimental results show that the system has many advantages such as non-touching, high speed, higher precision and strong anti-jamming etc. Key words: Bearing, Machine Vision, Image Processing 1 Introduction Machine vision inspection generally refers to the process of using a vision system to determine the deviation of a product from a given set of standard requirements[1]. Visual inspection technology has outstanding advantages such as non-contact, high speed, appropriate accuracy, and strong anti-interference ability on site, which can well meet the needs of modern manufacturing industry. In 1984, the total sales of industrial vision systems in Western Europe reached US$5.89 million, and by 1989 it reached nearly US$43.2 million; in the United States, the total sales of vision systems reached US$60 million in 1994, and by 1996 it reached nearly US$700 million[2]. At present, the total global machine vision market has reached US$6 billion to US$7 billion, and is growing at an annual rate of 8.8%[3]. With the development of manufacturing industry, visual inspection technology shows broad application prospects. The outer diameter of precision bearings is one of the important inspection parameters of bearings. In order to improve the grade of bearing products and increase the added value of products, accurate measurement of bearing outer diameter parameters is an essential step in controlling the quality of bearing products. In actual production, instruments such as vernier calipers, micrometers and coordinate measuring machines (CMM) are generally used for measurement. These measurement methods are only suitable for offline manual contact inspection, and their inspection accuracy and precision are often related to the operator's experience and work attitude. These methods are not only costly but also inefficient and unsuitable for 100% online inspection. In order to adapt to the characteristics of large production batches, strict quality requirements, and heavy inspection tasks in the bearing manufacturing industry, as well as the requirements of automated assembly line operation and achieving the inspection target of "zero scrap rate", this paper adopts computer-based vision inspection and image processing technology to design a bearing outer diameter inspection system based on machine vision. 2. Composition of the inspection system: A typical vision system generally includes an image acquisition system, an image processing part, a communication and I/O part, as well as input and output and actuators. The image acquisition system generally consists of a light source, a lens, a camera, an image acquisition card, etc. The image processing part includes image processing hardware and image processing software. Hardware selection for this vision inspection system: (1) Selection of light source: The light source has a great influence on the quality of target imaging. It directly affects the quality of input data and at least 30% of the application effect. This system uses an incandescent lamp that is easy to obtain, inexpensive, and easy to operate as the light source. The lighting system adopts backlighting. Backlighting is when the object to be measured is placed between the light source and the camera. Its advantage is that it can obtain high-contrast images. (2) Selection of image acquisition card: This system uses the Tenmin 10MOONS SDK2000 high-quality PCI video card. The display resolution can reach 640×480, 24-bit true color. Dynamic captured images are stored as static images, providing multiple storage formats such as BMP and JPG. (3) Selection of camera: Cameras are divided into CCD cameras and CMOS cameras according to their image sensors. Currently, CMOS is still inferior to CCD in terms of noise reduction and sensitivity. This system uses the Tenmin 10MOONS CCD lens. The highest resolution reaches 756×576 pixels, 24-bit true color. 3 Detection principle and process of this system: Under the condition of a certain measurement field of view, the most direct way to improve the measurement accuracy of the optical measurement system is to increase the resolution of the CCD camera, that is, to increase the number of pixels. However, the cost of increasing the hardware resolution is quite expensive and limited. This system does not improve the measurement accuracy by increasing the hardware resolution, but researches and develops an edge extraction algorithm with higher accuracy, that is, a sub-pixel algorithm. The diameter detection process is as follows: 3.1 Acquisition of the target image: The workpiece to be measured, the light source, and the CCD camera are stably fixed on the worktable, and the distance between them can be adjusted. Before the entire system works, the positional relationship between the workpiece to be measured and the CCD image-sensitive surface, as well as the focal length of the lens, need to be adjusted to ensure that the workpiece to be measured is clearly imaged on the CCD. The central axis of the workpiece to be measured is parallel to the CCD image-sensitive surface. The CCD camera outputs a standard PAL video signal, while the computer can only process digital signals. Therefore, the system uses an image acquisition card to filter and convert the video signal (analog signal) to A/D, and saves the result in memory. The application software in the microcomputer (10Moons SDK-2000 Video Capture 5.0) reads the digital signal and saves it as an image file (BMP format file). The image acquired in this experiment is shown in Figure 1 below. The image attributes are 720×576 pixels, 24-bit true color. 3.2 Image preprocessing: The preprocessing of the detection image includes filtering and denoising of the detection image, converting the color image to a grayscale image, etc. The main causes of noise in images are sampling, quantization, transmission, and environmental disturbances during image acquisition, including additive noise and impulse noise or salt-and-pepper noise [4]. In the implementation of computer vision detection technology, due to the requirements of detection speed and algorithm implementation difficulty, it is often necessary to convert color images into grayscale images for processing. Based on commonly used grayscale algorithms, this paper analyzes the characteristics of color detection images and studies a grayscale algorithm that is simple to calculate, can visually distinguish the grayscale of the background and the object, and can separate the grayscale levels of the object and the background in the grayscale image for subsequent computer processing. The image after grayscale processing in this experiment is shown in Figure 2. Image smoothing technology refers to a class of technologies that can smooth additive, impulse, and other noises to reduce the impact of noise on the image. Commonly used smoothing technologies include linear smoothing filtering, nonlinear filtering, and mathematical morphology filtering. This paper uses the median filter, which is the most representative nonlinear filter. The main reason is that it has a good filtering effect on certain specific random noises, and it has much less blurring effect when processing images than a linear smoothing filter of the same size. This paper uses a template to perform median filtering on the detected image, and the processed image is shown in Figure 3. The processing results show that median filtering can basically eliminate noise in the image, thus providing a "clean" image for subsequent processing. 3.3 Edge Tracking: Edge detection is a key problem in image processing, computer vision, and pattern recognition. Commonly used edge detection operators include Sobel, Prewitt, Laplacian, and LoG operators. Sobel and Prewitt operators can provide good and accurate edge information, but they also detect many false edges, resulting in low edge localization accuracy. In actual image processing, the Laplacian operator is generally not used directly because it is very sensitive to noise. Laplacian processing of images will produce double edge amplitude, making image segmentation difficult; moreover, since the Laplacian operator is an isotropic operator, it cannot detect the direction of the edge. This paper chooses the LoG operator as the integer-pixel-level edge localization function because it has the functions of both noise filtering and edge detection for the detected image. The image after edge tracking by the LoG operator in this experiment is shown in Figure 4. 3.4 Binarization: In computer vision inspection of mechanical parts, image binarization is a prerequisite for pattern recognition and geometric parameter data processing. A binary image is an image with only black and white values. Image binarization can be seen as a form of image compression. In most cases, this compression process will lose some image details. Therefore, in the binarization process, we should try to minimize the loss of image details relevant to practical applications. In this study, the edge region of the image target is the most important information in image processing; that is, in the binarization process, we must ensure that the target edge information is preserved as much as possible. In this paper, considering the trade-off between the accuracy and processing speed of binarization, an image binarization processing method is presented, which uses the image weighted standard deviation as a threshold and combines gradient and Laplacian edge detection. The image after binarization in this experiment is shown in Figure 5. 3.5 Subpixel Edge Localization: Generally, the accuracy of subpixel algorithms is verified by simulation. That is, several targets with known precise locations are simulated in the image, and then the corresponding subpixel algorithm is used to locate the targets. Finally, the accuracy of the algorithm is calculated based on statistical principles. This system uses a high-precision spatial moment edge sub-pixel localization algorithm. [align=center] Figure 1: Original true color image Figure 2: Image after grayscale processing Figure 3: Image after 3×3 median filtering Figure 4: Edge tracking using LoG operator Figure 5: Image after binarization[/align] 3.6 Diameter calculation: First, calculate the number of pixels of the target in the diameter direction, and calculate the diameter D using formula (1-1): D=N×dx (1-1), where, is the number of pixels in the diameter direction; dx is the object surface resolution, that is, the size of one pixel in the horizontal direction. The diameter detection method used in this system can detect the diameter value on each cross section of the bearing, so that the size of the bearing can be accurately determined. The detection results in section 5.2 below list the diameter values ​​and other parameters of each of the 18 cross sections of the bearing. 4 Calibration method in this paper: The commonly used calibration methods can be divided into traditional camera calibration methods and camera self-calibration methods. This paper studies the problem of detecting the parameters of mechanical parts. High detection accuracy is required, and the relative positions of the camera and the part in the detection system generally remain constant. Therefore, a traditional calibration method was chosen for the camera. A standard bearing was selected as the calibration reference. A standard bearing refers to a qualified product provided by the user, whose outer diameter is known in advance. The light source, CCD, and computer used in the calibration of the visual inspection system in this paper are the same as those used in the detection system; moreover, the position of the calibration reference, i.e., the standard bearing (relative to the light source and relative to the CCD), is the same as the detection position of the bearing being inspected in the detection system. This calibration method features high accuracy, a simple calibration process, low cost, and ease of implementation. The calibration process does not require calculating the camera's 13 internal and external parameters; it only requires calculating the camera's object plane resolution dx (dx = D/N, where D is the diameter of the calibration reference and N is the number of pixels between the two edges of the target in the horizontal direction). Furthermore, the calibration object used is the object being inspected itself, ensuring that the optical parameters at each position on the detection surface of the object are consistent with the parameters obtained from calibration. 5 Measurement Results Analysis and Conclusion: This system uses Visual C++.Net to develop a set of image processing software for measuring the outer diameter of bearings [5][6]. The main functions include: opening and saving images, filtering and denoising, edge detection and tracking, grayscale conversion, edge sub-pixel positioning, bearing outer diameter detection, etc. After the above processing, the bearing outer diameter is measured as follows: 5.1 Calibration results: 5.2 Measurement results: According to the above measurement results, for a bearing with an actual outer diameter of 47.800mm, the measured outer diameter value is 47.804+0.052/-0.039mm. The dispersion of the average diameter is 0.091mm. It can be seen that the bearing outer diameter machine vision detection system studied in this paper has achieved high detection accuracy. It realizes non-contact, online real-time measurement of bearing outer diameter and achieves the goal of 100% product inspection by modern enterprises, improving detection efficiency and accuracy. References: [1] BGBatchelor and DWBraggins. Commercial vision systems in computer vision [J]. Theory and Industrial Applications, 1992: 405-452 [2] Braggins, DW Image processing for industrial applications in Europe [J]. Proceedings of the 5th International Conference on Robot Vision and Sensory Control, 1990: 13-23. [3] Liu Zhaoni, Lei Zhenshan, Meng Li, Hou Rongtao. Gear measurement method using machine vision technology [J]. Tool Technology, 2004, 38(9): 133-135. [4] Gonzalez. Digital Image Processing (Second Edition) [M]. Beijing: Electronic Industry Press, 2003. [5] Zhong Guangzhi, Lu Jun, Liu Weirong. Visual C++.NET Digital Image Processing Examples and Analysis [M]. Beijing: Tsinghua University Press, 2003. [6] Yang Min. Research on Engine Valve Stem Diameter and Roundness Detection Based on Machine Vision [D]. South China University of Technology Library: Reference Room, 2004.