Different noise types may be introduced into light field data during acquisition, transmission and processing, due to low light conditions, corrupted pixel values, the statistical nature of electromagnetic waves, faulty memory space in storage, and hardware damage. Data noise seriously affects the accuracy of subsequent light field imaging techniques like depth estimation and post-capture refocusing. Therefore, light field denoising is important in light field imaging. 4D dual fan, 4D hypercone, and 4D hyperfan are classical filters designed based on the structure of the light field spectrum support. These filters can achieve light field denoising by passing the light field signal on the spectrum support while eliminating a significant amount of noise energy that lies outside the spectrum support. However, the noise suppression effect of the filter is poor for the noise located on the spectrum support, and the aliasing effect on the spectrum support also seriously affects the denoising quality. To further improve the denoising effect of the filters, we explore the structural characteristics of the spectrum support conducive to denoising effect improvement for determining the light field reparameterization. Both the quantitative indicator and the visual effect of the denoising results are improved by reparameterizing the noisy light field properly before denoising. Moreover, the idea of determining the light field reparameterization based on structure characteristics of the spectrum support provides a new perspective for improving the processing effect of light field data.

We consider two structural characteristics of the light field spectrum support, including the symmetry degree and the angle between two boundaries of the spectrum support. Analysis of how the two characteristics affect the denoising effect shows that the smaller angle and higher symmetry degree of the spectrum support are beneficial for enhancing the denoising effect. Based on the two structural characteristics, the concentration concept of the light field spectrum support for light field denoising is proposed, and the concentration degree metric function is designed. We can obtain the distance between two planes of reparameterized light field which is more favorable for denoising by minimizing this metric function, and the denoising effect can be improved by reparameterizing the noisy light field at this distance before applying filters.

Denoising experiments are conducted on both synthetic and real light field data. For synthetic data (12 HCI light field data), the PSNR and SSIM of the denoising results are both improved by introducing proper reparameterization compared with direct denoising under the same noise level and noise type (Table 3). Under different noise levels and noise types, the denoising results PSNR and SSIM obtained by introducing proper reparameterization before denoising are also improved (Figs. 11 and 12). Furthermore, after zooming in on the smooth area of the Pens light field, the direct denoising method still leaves obvious noise, while the reparameterization method eliminates the noise more effectively (Fig. 13). Zooming in on the area of Cotton light field where there is edge and reflection information reveals that the direct denoising method leaves obvious noise with the loss of edge information and reflection information, while the reparameterization method removes more noise and preserves better edge and reflection information (Fig. 14). For real light field data, the reparameterization method can provide better denoising effect compared with the linear shift-invariant filter, which is another type of filter based on the spectrum structure (Table 4).

Our paper considers two structural characteristics of the light field spectrum support, including the symmetry degree and the angle between two boundaries of the spectrum support. We analyze how the two characteristics affect the denoising performance, and propose the concept of the concentration degree of the spectrum support for light field denoising and its corresponding metric function. The distance between the two planes of the reparameterized light field is obtained by minimizing the metric function, and the reparameterization is introduced to improve the light field denoising effect. The synthetic light field experiments show that by minimizing the metric function of concentration degree, the spectrum support of the light field data becomes more concentrated. For the classical 4D hyperfan, 4D dual fan, and 4D hypercone filters, the introduction of proper light field reparameterization can improve the denoising quality of light field data, and PSNR and SSIM are increased under different noise levels and noise types. Additionally, more edge and reflection information can be preserved with more noise removal. The real light field experiments show that compared with the linear shift-invariant filter, classical filters with reparameterization yield better denoising results in both PSNR and SSIM values. In addition to light field denoising, exploration of the target spectrum structural characteristics of other computational imaging tasks, and application of corresponding reasonable reparameterization of light field data, the proposed idea of introducing reparameterization before light field data processing can be beneficial for other specific computational imaging tasks.