Aiming at the common problem of end face defects in the production process of bearing manufacturers and the current situation of manual visual inspection, an online detection method of bearing ring end face defects based on machine vision is proposed. First, the ferrule image is preprocessed for edge detection, and the four-connected domain is used to locate the end face area of ​​the ferrule; secondly, the least squares method is used to fit the end face contour to identify the shape defects, and polar coordinate transformation is used to stretch the annular end face of the ferrule into a Rectangle, the Sauvola local binarization algorithm is used to segment the rectangle image, and the defect image is converted back to a ring image through the coordinate system inverse transformation and bilinear interpolation method; finally, the defect identification and classification are completed according to the image features of the extracted defects. . Field tests show that the overall recognition accuracy of the ferrule end face detection system is 98.6%. Bearings are an indispensable basic component to ensure the rotation accuracy of mechanical equipment. After the bearing ring is surface-ground, its end face may still have appearance defects such as forging waste, large and small edges, bumps, car waste, abrasion, black skin, etc. . If there is a defect on the end face of the ferrule, it will be used as the positioning surface of the post-processing station such as the outer centerless grinding, which will inevitably affect the machining accuracy and the rotation accuracy of the bearing, which may cause noise and vibration during the use of the bearing, thereby accelerating the wear and even causing the machine. Fault. On the other hand, the removal of defective ferrules after entering the subsequent process or the recall of finished products after entering the market will bring great waste of material and labor costs to the enterprise. Therefore, defective products must be removed after surface grinding to avoid flow into subsequent processes. At present, most enterprises still rely on the naked eye and subjective experience of quality inspectors to identify and judge the ferrules. The quality inspection results are easily affected by human factors, the inspection standards are difficult to keep consistent, the stability is poor, and it is easy to miss inspections. Machine vision has the advantages of high precision, high efficiency, and good real-time performance, and is an effective method to replace manual detection. For example, in literature [2], the improved Otsu method is used for threshold processing, and the eight-connected domain marker recognition technology is used to realize the bearing end face. Non-contact detection; Reference [3] solves the problem that the changing light intensity of the air bearing surface affects the image acquisition by using the texture unit. On the basis of the above research, this paper proposes an online visual detection method for bearing ring end face defects, using four-connected domain, seed filling algorithm to locate the detection area, Sauvola local binarization algorithm for image segmentation, and appearance defects based on multi-features Identification methods identify defects. 1. Light source selection and detection area positioning 1.1 Light source selection The light source is an indispensable part of the visual inspection system, which is directly related to the imaging quality. A good lighting method can highlight the characteristics of the target area and reduce the workload of image processing. The imaging surface of the end face defect detection is a circular metal end face, which has a certain specular emission effect, and the size of the tested ferrule spans a large span, so the spherical integral diffuse reflection shadowless lighting method is used, which has a large irradiation area, concentrated and uniform light, and no A specular reflection is formed, as shown in Figure 1. Figure 1 Spherical integral diffuse reflection shadowless illumination1.2 Image preprocessing Noise and interference during image capture will reduce image quality and increase the difficulty of subsequent edge detection and image segmentation. Therefore, it is necessary to The original image is subjected to certain preprocessing to eliminate noise and interference in the image. The comparison effect of mean filter, Gaussian filter, and median filter is shown in Figure 2. Because the background area is outside the annular area of ​​the end face, the gray value is low, the light-dark contrast of the annular area after the mean filter processing becomes weaker, and the image becomes blurred; the edge part of the end face after the median filter processing is affected by the black area, and the edge details are lost. ; and Gaussian filtering, due to the characteristics of weighted average, can well preserve the details of the ring and its edges while removing noise. In view of the large contrast between the end face area of ​​the ferrule and the background area, and the need to detect small defects, a filter window with a size of 3 × 3 and a standard deviation of 1 is selected for Gaussian filtering. Fig.2 Effect diagram by different filtering methods1.3 Edge detection Edge detection is to determine the position of the image edge that needs to be identified by identifying the part of the image where the brightness changes significantly. The grayscale mutation in the region reflects the important changes of the image and is an extremely valuable image feature. For this study, edge detection can well complete the division of the ferrule end face area and the background area, making it easier to locate the detection area and prepare for the next step. Common edge detection operators include Canny operator, Sobel operator and Laplacian operator. The Sobel operator has a strong ability to adapt to the edge of noise and grayscale gradient, but it has the function of smoothing the image, which is suitable for occasions that do not require high accuracy; the Laplacian operator has poor anti-interference ability to noise, which will cause invalid pixels to be removed. regarded as edge points, but will accentuate the contrast of edges,It is suitable for image sharpening scenes; Canny operator has the characteristics of low error rate, strong positioning ability, and single edge pixel response, and is called the best edge detector. The comparison of the three operators is shown in Figure 3. The Canny operator has stronger anti-noise interference ability, stronger edge positioning ability, and can detect real weak edges.