Research On The Key Technology Of Wheel Hub Casting Defect Detection

With the rapid development of information technology,the automatic detection and grading of casting defects of wheel hubs can improve the quality and efficiency in production.This paper mainly studies the object detection and object segmentation technology of wheel hub casting defects to reduce the missing detection and excessive detection in hub defect detection,and to grade the defects.The accuracy of traditional object detection algorithm depends on parameters of manual adjustment,and missing detection rate and excessive detection rate are both high,which can not meet the needs of actual production.Faster R-CNN in the R-CNN series object detection models based on convolutional neural network(CNN)is applied to the object detection of hub defects,which missing detection rate and excessive detection rate are both lower,and there is no need for manual intervention.According to the characteristics of hub defects,this paper proposes a new anchor box setting proposal,and adopts the basic network with stronger ability and multi-scale information fusion with multiple convolutional layers.Our experiment proves that the improved model has a lower missing detection rate.The main task of hub defect object segmentation is extracting more accurate defect contour,and calculating defect area and extension length to grade the defect.Due to the unsatisfactory result of the traditional algorithm in complex scenes,this paper uses Fully Convolutional Network(FCN)to segment hub defect object,and using lower level information to assist the training method of up-sampling can obtain more accurate segmentation results.To further improve the effect of segmentation,a mask branch is extended on the basis of Faster R-CNN,and obtaining Mask R-CNN,which can not only locate and segment the defect object,but also eliminate the false object detection result by segmentation result,moreover,the defect area and the defect extension length are calculated by the object segmentation results to achieve the grading of defect.Experiments show that compared with Faster R-CNN,Mask R-CNN has lowerexcessive detection rate and better object detection effect,and compared with FCN,the segmentation result of Mask R-CNN is more accurate,which average IoU is increased by 8.7963%.Moreover,the error of defect area and extension length calculated from the segmentation result of Mask R-CNN is also relatively low,and the result can meet the requirement of defect grading.Finally,Our realize an automatic detection system of hub defects based on the object detection and segmentation algorithm of wheel hub defects in this paper.