The paper presents a method for processing industrial tomographic images based on a combination of blind denoising using the Noise2Self principle and subpixel super-resolution. A key advantage of the developed approach is the ability to train without the use of reference “clean” images, which is especially critical for industrial applications where obtaining paired data is challenging. Experimental studies conducted on 400 sets of projection data from industrial objects demonstrated an improvement in peak signal-to-noise ratio by 8.4 dB and structural similarity index measure by 0.16, while preserving critically important technological defects. The proposed method is adapted for operation in multi-material scenarios, including metals, composites and polymers, and supports the most common industrial data formats. It is noted that integrating a metal artifact detector into the preprocessing block helped to minimize the distortions when working with multi-component materials. The paper also discusses potential directions for further development of the method, including the implementation of a 3D pipeline, the application of transfer learning for equipment-specific adaptation and the creation of a hybrid physics-neural network model. Particular attention is given to the prospects of deploying and integrating the solution into computer tomography systems.