Automating defect visibility assessment in industrial radiography and realistic film noise modelling based on experimental investigations

Abstract

Radiography is one of the main methods within the industrial field of Non-Destructive Evaluation (NDE). It is a projection imaging technique relating the subsurface information of the structure under investigation to a greyscale valued image. In order to ensure operational reliability, inspection procedures need to be qualified before the actual evaluation takes place in practice. Alongside experiments, this qualification process often involves computational tools to model various possible test setups incorporating artificial defects, producing a simulated image. Then the simulated image is assessed, which is traditionally performed subjectively by an operator deciding on the visibility of an artificial defect. If the detectability of that defect is increased, then the underlying parameters of the considered test procedure are preferred. Two crucial aspects of the stated qualification process are addressed within this thesis. Firstly, the presented work advances the simulated qualification approach to a more holistic one by not only modelling the test setup and simulating the resulting image but also by assessing the simulated image automatically. A Model Observer approach is chosen for automating the interpretation of the simulated images. This approach is based on the so-called Channelized Hotelling Observer (CHO), which was originally established in the medical field but is now transferred to NDE. Validation of the new assessment approach is done by comparing experimental psychophysical data of human image assessment to the performance of the CHO in assessing simulated radiographic images. The new method outperforms other state-of-the-art visibility models currently used in industry. Secondly, the correct simulation of the artificial radiograph is crucial for an accurate computational qualification of the test procedures. An important part of the radiograph is the underlying noise, which has a great influence on the image quality. Industrial radiography still mainly relies on films as detectors, since these are reliable and well established over decades. These films show a typical film noise granularity, which usually degrades the defect visibility. Therefore, it is crucial to incorporate realistic film noise in the simulation by a sophisticated film noise model. Currently, the industry considers only white Gaussian noise, which is unrealistic. A new noise model based on experimental investigations of real noise on film samples is presented. It is shown that very realistic artificial noise can be generated by extracting and utilising the spectral characteristics of the real noise. The method generates new and unique noise samples by randomising the phase spectrum but keeping the original spatial frequency spectrum. Furthermore, the method allows the generation of noise with local, space depending spatial frequency spectra while ensuring pixel correlation throughout the image. Various validation studies show that the noise generated by this new method is visually and statistically indistinguishable from the original noise. Additionally, a new optical setup for digitising radiographic films on the microscale is presented. This setup is based on a digital camera and image processing procedure enabling the extraction of the film noise spectra that can then be used to generate film noise realistically. Homogenously exposed and developed films are investigated by the new setup in order to generate a database of spectral properties necessary for the generation of accurate film noise. Furthermore, the new setup can efficiently classify the quality of film systems without relying on time-consuming microdensitometer measurements, which is an important secondary benefit of this new setup.