Welding Defect Inspection And Recognition Base On X-ray Images

Due to the interference of the external environment, weld seam often inevitablyappear some defects such as fusion, crack, pore, slag inclusion and so on. X-rayinspection has the advantage of stabilization and high resolution, thus it’s widely usedin industrial production. Recently, with the development of computer imagingtechniques, defects detection based on X-ray gets more and more attention. But, onthe one hand, X-ray imaging system itself limits the image quality, and the image hasthe low signal noise ratio (SNR). On the other hand, weld image has the feature ofhigh complexity, easy to distortion, small size etc. These shortcomings limit thepractical application for promotion.This paper design and optimize defect detection algorithms including imagepreprocessing, segmentation, welding defects extraction and classification algorithm.Through the theoretical research and experimental analysis demonstrate that thesystem has the superiority of discriminant accuracy and stability, and can widely usein the weld defect detection. The main process is listed as below:(1) In the actual production application, weld seam distribution usually hascertain complexity, in the process of X-ray transmission, it has relatively significantinterference, that causing subsequent difficulty of defect recognition. Under theframework of multi-scale contour mathematical morphological filtering， thepreprocessing of complex welding image is discussed and good filtering effect isobtained．(2) Due to the parts size, weld position in X ray images also exist in thecorresponding deviation, traditional detection method cannot adapt to the randomfluctuations, and will ultimately affect the adaptability of the detection system.Through the analysis of X-ray weld image, the supervision of transition regionextraction and threshold segmentation algorithm is proposed. The new descriptor isadopted firstly–the fusion of local fuzzy entropy and fuzzy variance to improve theinsufficiency of traditional transition region descriptor, which uses local fuzzy entropyand local fuzzy variance respectively reflect the gray level changes in the neighborhood window. Due to traditional transition region extraction method directlyoperation to the original image, the prior knowledge is used to overcome the shortageof the traditional method.Secondly, in view of the traditional extraction of transition area failed to considerthe shape of the neighborhood window, which can’t accurately depict the gray levelchange in the neighborhood, a new descriptor based on the minimum gray differenceis proposed to represent transition gray level changes. Due to traditional transitionregion extraction method fail to consider the characteristics of human visualperception, the unsupervised transition area extraction method is constructed. Aftergetting the segmentation binary map, the low gray level pixels are looked up todetermine whether to find the region of interest. Experiment Results show that thenew method can extract the clear weld seam area, and it is very close to people’svision.(3) Defect extraction is the important link in X-ray imaging automatically detectsdetection. Because the size of the internal defects are relatively small and easy underthe influence of uneven welding composition of internal organization, the design andoptimization segmentation algorithms face serious difficulty. In the help of artificialneural network, through the analysis of characteristics of X-ray weld seam image,edge detection method base on LVQ neural network is designed to remove thebackground from the original image. firstly the representative characteristic vector isdesigned to train the neural network, and then each pixel’s vector is extracted in theimage which as the input of neural network, finally binary image denoising algorithmbasd on mathematical morphology is designed. The denoising image can truly reflectthe shape and size of the defect, it is the ideal image processing tools.(4) Defect classification is the basis of weld seam quality assessment, base on thecharacteristic of X-ray weld defects image, five characteristic vector is chosen as themost optimal combination. These parameters are as the input of subsequent SVMclassification method to distinguish the nature of the defect. According to the characteristics of welding defects, classifier based on support vector machine with thehelp of binary tree is designed. Through the experiment resulst prove that theproposed method is correct and effective.