Abstract
The initial visual encoding results in substantial information loss, arising primarily from optical blur and from spatial and spectral sampling by the cone mosaic. Previous research has investigated this information loss using Bayesian image reconstruction applied to the cone excitations, exploiting a sparse-coding prior for natural images. Here we study the effect of using a more expressive natural image prior. Specifically, we adopt a recently developed method for sampling from the image prior implicit in a denoising convolutional neural network, and combine this with constraints from a linear model of the initial visual encoding. The network was trained to denoise images from the ImageNet dataset that were corrupted with additive Gaussian noise. The linear measurement model was obtained through the open-source ISETBio vision model. Input images were 128 by 128 pixels, and simulations were performed for 1-by-1 degree patches of retina. We compared reconstruction results obtained with the denoiser prior and the previously used sparse-coding prior. For both priors, image reconstructions from the fovea were close to veridical. In an analysis that used the accuracy of the reconstructed images to examine how the information about natural images varies with the allocation of L, M, and S cones in the foveal mosaic, the two priors led to similar conclusions. This supports the idea that this analysis is robust to the particular choice of prior. In visualizations of reconstructions from peripheral retinal mosaics, key qualitative features such as a decrease in spatial detail and reduction in chromatic contrast were consistent across the two priors. Nonetheless, the reconstructions from the two priors are visually distinct: The denoiser prior produced images with sharper edges and more continuous contours, which generally appear more “natural”. These differences could be important for some quantitative applications, such as the development of perceptual image quality metrics.