It is still difficult to identify and detect hidden faults in multi-layered composites like fiber-metal laminates (FML), even using advanced Xray Computer Tomography (CT). For example, an impact damage is nearly invisible by a frontal Xray projection, although, the deformation can be seen and detected manually by hand perception. Things are getting more worse if a portable low-cost Xray radiography or semi-tomography machine is used (called Low-Q measuring device), as introduced and described in this work. The Xray equipment consists of a low-cost Xray source for dental diagnostics and an Xray detector consisting of a conventional medical Xray converter and amplifier foil (Fine 100) backside imaged by a commercially available CMOS monochrome image sensor (back illuminated Sony IMX290 2M pixel sensor) and a simple two-lens optics. The optical distortion introduced by the optics increases with increasing distance from the center of the image ("barrel distortion") and must be corrected, at least for CT 3D volume reconstruction.
The measured Xray images pose increased spatially equally distributed gaussian noise (compared with high-quality flat panel detectors) and more important randomly located "popcorn" shot noise by avalanche effects in pixels and pixel clusters (islands) due to Xray radiation exposure (back-illuminated sensors are very sensitive for this noise). The gaussian noise can be reduced by averaging, the shot noise is removed by using multiple images recorded in series and an automated pixel replacement algorithm. The shot noise is a seed threshold phenomena, i.e., the location and number of white pixels changes from image to image, therefore allowing the replacement of white pixels in one image from unaltered pixels from another image.
After the image preprocessing and filtering stages, damage and material faults are identified by a pixel anomaly detector, basically an advanced Convolutional Neural Network (CNN) and region-proposal R-CNN models. Training of pixel classifiers can use a few images only because each pixel region is a sample instance. R-CNN models require an extended sample data base, which cannot be acquired only by physical measurements. The training and test data set will always limited by a limited number of specimens, e.g., with impact damages, and a limited variance in material and damage parameters (e.g., location). For this reason, the data set is extended by synthetic data augmentation using Xray simulation. In contrast to other wave measurement principles like Guided Ultarsonic Waves (GUW), Xray images can be simulated with high accuracy (compared with physically measured images). We are using the gVirtualXray software library [GVX23;VID21] performing Xray image simulation by using GPU processing only and raytracing. GvirtualXray is proven for its suitability to produce accurate images as long as diffraction and reflection of Xrays are neglected. In addition, a novel few projection hidden-damage detection methodology is introduced that can be used in-field with a portable Xray machine as described above. Preliminary results show a high damage detection rate for a wide range of materials.
## References
[GVX23]: gvirtualxray, https://gvirtualxray.fpvidal.net, accessed on-line on 24.1.2023
[VID21]: F. P. Vidal, Introduction to X-ray simulation on GPU using gVirtualXRay, In Workshop on Image-Based Simulation for Industry 2021 (IBSim-4i 2020), London, UK, October, 2021
