Analysis and Correction of Micro CT Images for 3D Reconstruction

In Micro Computed Tomography (Micro-CT), which is a non-destructive technique used to reveal the 3D profile of an object, image slices are often the result of the procedure. Subsequently, to retrieve the original object, 3D reconstruction is typically carried out, where by these slices are often combined in one form or another, using various techniques such as photogrammetry to eventually reproduce the object. Due to its high-spatial scale, the microtomography procedure is usually prone to some misalignment issues that are induced by the mechanical operation of the system itself and result in errors in the image slices produced. This paper focuses particularly on analyzing the image slices produced and attempts to detect and correct the abnormalities by implementing a detection and correction pipeline. A total of1441 electronic circuit-board image slice dataset were used in the experiment. We adopted a deep-learning method, by building and training a custom cGAN based Pix2Pix network. The network was trained in a supervised setting, using consecutive image pairs as sources and targets respectively. It was designed to implicitly learn the transform for each consecutive image pair, such that after training, a source image could be fed to the model to generate a predicted target-PRIME image that would serve as the correction for a detected abnormal target image flagged based on a pre-selected threshold score value. Using the selected threshold score, 48 abnormal images were eventually successfully detected and corrected.