Image Super-resolution Algorithm for Fan Beam CT Based on Denoising Diffusion Probabilistic Models

With the development of advanced manufacturing industries, such as the emergence of 3D printing technology, the structure of metal parts is becoming increasingly complex. However, traditional non-destructive testing techniques are powerless for measuring internal dimensions with cavity structures. Fan beam CT technology collects multi angle projections of the object to be measured through a linear detector, and combines reconstruction algorithms to observe a certain section, thereby achieving size measurement and defect detection. The micro-motion technology of detectors aims to merge multiple discrete data points that are less than one pixel apart by making small movements of the detector in the array direction, in order to obtain images with higher resolution than the detector. However, the micro-motion technology of the detector will prolong the CT detection time. To address this issue, this paper proposes a fan beam CT super-resolution technique based on conditional denoising diffusion model[1]. This method uses high-resolution images obtained by detector micro-motion technology as the starting point of the diffusion model, and low resolution images obtained by detector micro motion technology as conditions. Through forward add-noising and backward denoising processes, higher resolution fan beam CT images can be obtained without the need for detector micro-motion technology. This method overcomes the problem of excessive smoothness in images caused by regression-based deep learning methods, and has good preservation of small defects in images. In human eye evaluation tests, the image effects generated by this method far exceed those based on CNN[2], Transformer[3], GAN[4], which has a great promoting effect on improving CT detection efficiency.