Overcoming image degradation in fiber-bundle tomographic combustion measurements via generative adversarial super-resolution of raw projections

Fiber-bundle-based tomography for combustion diagnostics suffers from severe image degradation due to signal attenuation and optical blur, which compromises three-dimensional (3D) reconstruction accuracy. Existing super-resolution (SR) methods focus on simulated volumetric data rather than addressing the root cause: poor raw projection quality. This paper proposes an ESRGAN-based framework to enhance raw two-dimensional (2D) projection images acquired by a nine-view bifurcated fiber bundle before reconstruction. A high-speed camera provides paired high-resolution references for supervised training. Experiments are conducted under laminar and turbulent ethylene flames with attenuation, blur, and combined degradation modes. Quantitative results show that under combined degradation, the SR images consistently outperform bicubic interpolation, yielding PSNR improvements of 2.15 dB (laminar) and 1.84 dB (turbulent), while reducing LPIPS by 44.77% and 42.18%, respectively. When used for tomographic reconstruction, the effective voxel count increases by 15.9 percentage points (laminar) and 9.0 percentage points (turbulent). Reconstructed slices exhibit sharper flame boundaries and preserved fine-scale structures. This source-level enhancement offers a more effective pathway toward high-fidelity fiber-bundle combustion tomography.