Detection Of Porosity Defect Of Casting DR Images Based On YOLACT Deep Network

Defect detection is one of the important steps to ensure the quality of industrial castings,which can reduce the safety risks that caused by defects.Digital X-ray imaging(Digital Radiography,DR)technology,as a commonly used industrial nondestructive testing technology,is an effective method for detecting defects in industrial castings.At present,the defect detection method of DR images in industry is mainly manual visual inspection.This method is not only time-consuming,labor-intensive and prone to fatigue,but also largely relies on the subjective factors of the inspectors,which is not conducive to the huge DR image data set.Therefore,it is of great significance to realize real-time detection of DR images defects by adopting automatic detection methods.The rapidly developing deep learning is widely used in various detection and recognition tasks.But the detection methods based on deep learning need data-driven.The currently published training data sets do not include the casting DR image defect data set.And the defect targets in the casting DR image are not prominent enough,which affects the detection and recognition of defects.In the process of detecting casting defects,it is hoped that real-time detection can be achieved.Therefore,this thesis takes DR images of castings(such as bolster and side frames)and mainly focuses on the detection of DR images for porosity defects.The following tasks are mainly completed:(1)Aiming at the problem that it is difficult to directly use the original DR image for defect detection and classification,this thesis integrates a variety of image enhancement methods to enhance the original DR image.Firstly,adjusting the window width/transform of the original DR image to highlight the general shape of the defects in the casting image.Then,using guided image filtering combined with fractional differentiation for image enhancement: guided image filtering has edge-preserving and smoothing effect on DR images,and the output image that is obtained by guided image filtering is subtracted from the input DR image to get the difference image.The difference image with the grayscale expansion and the output image are added to obtain an enhanced image;based on them,fractional differentiation is used to further enhance defects in the image.(2)In deep learning,data set is a crucial factor.There are almost no public data sets for Bolsters and side frames.This thesis uses the DR images of bolsters and side frames that provided by a company,and first perform the image enhancement on them;and then use the labeling software Labelme to label them to obtain the json file,convert each json file to COCO data set format,and finally form a training data set containing images and their annotation information.(3)In this thesis,a YOLACT(You Only Look At Coefficients)network is developed for defect detection of DR images.The training data set is input into the network,the network is trained and the network parameters are set to obtain the average accuracy of prediction box and mask for defects.DR images in the test set are tested to get the test result.The experimental results show that the defect detection method obtained by combining the image enhancement method with the YOLACT network can better detect the grade and location of the porous defects in DR images.The detection speed is fast,and the purpose of real-time detection can be achieved,which provides an auxiliary solution for the defect detection of industrial castings.