The pseudo-boiling line defines the transition of supercritical fluids between liquid-like and gas-like states. In its vicinity, small fluctuations in temperature or pressure lead to disproportionately large variations in density and compressibility. This phenomenon can be utilised to trigger flow instabilities by small perturbations and therefore enhance turbulence in micro-confined environments. Micro-PIV provides a valuable technique for visualising and quantifying fluid motion in microscale environments. However, its application under supercritical conditions faces several fundamental challenges. Strong refractive index gradients caused by density fluctuations introduce optical distortions that degrade image quality. Illumination constraints and limited optical access reduce signal-to-noise ratios, while high-speed imaging under low-light conditions amplifies sensor noise, making it difficult to extract reliable flow information. These limitations are particularly severe when investigating flows near the pseudo-boiling line, where optical turbulence and non-uniform scattering often obscure particle tracks. As a result, obtaining accurate, high-resolution velocity fields from micro-PIV measurements in supercritical systems remains an open challenge. The presented work introduces a novel data-driven method to enhance the quality of micro-PIV images acquired under low-light conditions in thermodynamically demanding regimes characterised by non-linear refractive index changes. First, an unsupervised denoising strategy is applied to raw images without the need for clean reference targets. Second, a despeckling network reconstructs particle images from distortions induced by highly non-linear refractive-index gradients along the optical path, using ground-truth images captured under undistorted conditions as training targets. Both networks were trained with a publicly available dataset of supercritical CO$_2$ generated by the authors. The developed framework highlights the capability of effectively restoring images corrupted by sensor artefacts and optical turbulence. This result further demonstrates the potential of combining deep-learning–based image reconstruction methods with experimental fluid mechanics under extreme conditions, and extending advanced flow diagnostics into regimes traditionally inaccessible to optical techniques.