Quantitative weld defect sizing using convolutional neural network-aided processing of RT images

Non-destructive confirmation of seamless welding is of critical importance in most applications and digital industrial radiography (DIR) is often the method of choice for internal flaw detection. DIR images often suffer from fogginess, limiting the inspection of flawed regions in online and quantitative applications. Much focus has therefore been put on denoising and image fog removal to yield better outcomes. One of the methods most widely used to improve the image is the fast and flexible denoising convolutional neural network (FFCN). This method has been shown to offer excellent image quality performance combined with fast execution and computing efficiency. In this study, the FFCN image processing technique is implemented and applied to radiographic images of welded objects. Enhancement of defect detection is achieved through sharpening of the image feature edges, leading to improved quantification in weld flaw sizing. The method is applied to the radiographic images using the weighted subtraction method. Experienced radiographers find that the weld defect detail is better visualised with output images from the FFCN algorithm compared to the original radiographs. Improvement in weld flaw size quantification is evaluated using test objects and the distance between the first two lines of the image quality indicator (IQI). The results show that the applied algorithm enhances the visualisation of internal defects and increases the detectability of fine fractures in the welded region. It is also found that, by selective image contrast enhancement near the flaw edges, flaw size quantification is improved significantly. The algorithm is found to be efficient, enabling online automated implementation on standard personal computers.