Diffuse light field imaging through thick scattering media

This is the authors' abstract. We don't add key points for this paper.

Scattering imaging through thick scattering media is often hindered by extremely low signal-to-noise ratios (SNRs) due to the ballistic-photon signal becomes negligible. To break through the super low SNR limit, we propose a diffuse light field imaging model (DLIM) to retrieve forward-scattered photons as signals. In this model, a diffuse source is constructed in an angular accumulated light field image by synthesizing the radiance sources captured from various view images. After that, the clear image could be solved analytically as a source term based on the diffusion equation by deconvoluting the diffuse Green function, which builds the diffuse light field convolution theorem. To demonstrate the physical process mathematically, we propose three-plane parametrization, which facilitates the derivation of DLIM from radiative transfer in each view image to diffuse approximation in the synthesized light field image. This is the first physically-aware scattering light field imaging model, extending the conventional light field imaging framework from free space into diffuse media. Extensive experiments confirm that the DLIM can reconstruct the target objects even when scattering light field images are reduced to random noise at extremely low SNRs. Compared to state-of-the-art scattering light field imaging methods like peplography, the proposed method outperforms by 1.70 dB/4.76 dB peak-signal-noise-ratio (PSNR) and 0.167/0.172 structural-similarity-index-measure (SSIM) higher, on average, for passive-luminous/self-luminous targets, respectively.