SFNS: Spatial-frequency image noise suppression for low-power industrial cone-beam computed tomography

With the widespread use of industrial computed tomography (CT) in semiconductor manufacturing, the noise in tomographic images from low-power radiation sources has become an increasingly prominent issue, which severely impacts the detection and processing tasks. In this paper, a spatial-frequency image noise suppression model (SFNS) is proposed to suppress noise while maintaining structural details in the image. A hybrid loss with spatial-frequency dual-domain information is designed to resolve spatial-domain limitations in optimization through complex-discrepancy-based spherical region separation criterion. To address the coupled noise characteristics in low-power CT systems, a stochastic noise degradation training strategy is designed to dynamically emulate fluctuations in real noise environments, and the network architecture integrates checkerboard sampling with cascaded residual stacks to enhance detail perception and reduce computational overhead. Experimental results demonstrate that SFNS achieves PSNR 32.14dB and SSIM 0.8896. Both quantitative metrics and qualitative evaluations validate the effectiveness of the proposed method in balancing structural fidelity and noise suppression.