Research On Defect Detection Of X-Ray Images Of Railway Casting Based On Mask R-CNN

With the increasing capacity and speed of Chinese railways,railway freight parts have higher inspection requirements.The manufacturing method of railway freight parts in China is mainly foundry.Due to many production procedures,casting technology and other influencing factors,casting internal defects are unavoidable,so defect detection is a key step in the production process of railway freight parts.Non-destructive testing can detect the defects of castings without damaging the castings.Industrial non-destructive testing generally uses ray detection technology,and Digital Radiography(DR)is the most widely used detection technology in industry.Most casting manufacturers use manual judgment methods and traditional image recognition methods to detect defects.The DR inspection speed is fast,about 3m / min.The manual judgment method has the disadvantages of large task volume,low judgment efficiency and greater influence by subjective factors;traditional The image recognition method can only perform single-task operations,such as defect positioning,defect identification,and other single tasks,cannot perform multi-task parallel detection,and cannot simultaneously classify and classify defects.Due to the emergence of Convolutional Neural Networks(CNN),deep learning has exploded,and target detection has entered the era of deep learning,from Regional Convolutional Neural Networks(R-CNN)to improved Fast R-CNN and Faster R-CNN to Mask R-CNN,the target detection step by step from the single task era to the multitask parallel era.In this paper,a defect detection method based on Mask R-CNN is proposed for the pain points of manual judgment and traditional image recognition methods,aiming to realize the classification and grading of casting ray DR image defects.The main research contents of this paper are as follows:(1)In view of the characteristics of scattering,noise and complex imaging of the workpiece ray DR image,image preprocessing methods such as window width/window transformation and image smoothing processing are performed on the original image.The smoothing of the ordinary filter will cause the loss of details and the accuracy of subsequent defect recognition will be reduced.Guided filter is used for image smoothing,then the smoothed image is differentiated from the original image to obtain a difference image,and the difference image and the smoothed image are added to perform image enhancement,makes the subsequent defect labeling proceed smoothly.(2)For deep learning networks,the training data set is a key factor in determining the performance of the model.Currently,there is no such training data set in the field of castings.Therefore,this article screens and processes the image data set provided by CRRC Qiqihaer Rolling Stock Company,uses labelme labeling software to mark defects,and finally form a standard database of 3000 images.(3)Construct a Mask R-CNN network,perform parameter setting and model adjustment on the network,and then use the labeled training data set to fine-tune the parameters in the pre-trained network model.Obtain the average detection accuracy and defect detection results of various types,and compare the defect identification results of YOLO v3.Experiment prove that the defect detection method using Mask R-CNN combined with the guided filtering enhancement method can better achieve the classification and classification of the defect detection of the DR image of the casting,and provides a solution for applying deep learning methods to the defect detection of industrial castings.