Research On Defect Detection Method Of Wire Braided Hose Based On Machine Vision

With the continuous progress of modern manufacturing technology,the quality inspection of industrial products is also developing in the direction of high precision and high efficiency.The existing production technology of steel wire braided pipe has been automated and upgraded,and traditional quality inspection methods have been unable to adapt to the production of steel wire braided pipe.Therefore,companies have begun to seek a quality inspection method based on machine vision.Machine vision inspection technology has the characteristics of high precision,high efficiency and non-contact.Target detection based on machine vision has become one of the important development directions of quality inspection technology.Based on the quality inspection requirements of a steel braided pipe proposed by an enterprise,this paper studies the following aspects in combination with machine inspection technology:(1)Through the study of image filtering,image segmentation and edge detection techniques in traditional image processing algorithms,the outline of the braided tube is extracted,and the diameter of the braided tube is sampled every other pixel,and the braid is estimated by calculating the variance of the diameter of the braided tube.The diameter of the tube fluctuates,find the area with large diameter fluctuation,and complete the detection of the diameter deviation defect of the braided tube.(2)In order to meet the requirements of real-time detection of steel wire braided tubes,this paper uses a deep separable convolutional layer to replace the ordinary convolutional layer,reducing the amount of calculation parameters of the YOLO network,and greatly improving the detection speed of YOLO;at the same time,through the convolution structure The improvement of,it can still maintain the original detection accuracy even when the amount of calculation parameters is reduced.Through this improvement,the detection speed of YOLO can be increased to three times,and the overall detection accuracy rate has been increased by about 1%.(3)Aiming at the problem that the detection rate of small-size defects in the YOLO network is not high,the method of improving the multi-scale network is adopted to increase the output feature layer of 4 times sampling in the multi-scale network,and the resolution of the output feature layer is increased.,And increase the pyramid feature fusion algorithm on the original basis to improve the information transfer between the feature layers.Experiments show that this improvement can increase the overall detection accuracy of YOLO by 7.89%(4)Adjust the a priori box and IoU calculation method in the YOLO prediction algorithm.1)The a priori frame size used in YOLO is suitable for detecting larger target objects.In the defect detection of steel wire braided pipes,the flaw size is generally small,and the original a priori frame size is easy to affect the fitting data distribution Therefore,this paper uses the K-mean algorithm to adjust the size of the prior frame.2)The IoU calculation method used in YOLO,when the prediction frame and the real frame do not overlap,it is easy to cause the problem that YOLO cannot be optimized during network training.Therefore,this paper uses the CIoU algorithm to improve the accuracy of YOLO during network training.Experiments show that the above improvements can increase the overall detection accuracy of YOLO by 4.51%.