The work is devoted to the perceptually meaningful images processing based on the neuromorphic autoencoder model inspired by the coding mechanisms of the human visual system. For adequate modeling of visual mechanisms, the neuromorphic model uses the most realistic representation of the input data in the form of a set of counts—a stream of photoreceptor firing events. The statistical model for counts is chosen in the form of a Poisson two-dimensional point process, the justification for which was obtained in our previous works. For this model a statistical description of the input data is proposed in the form of a sampling representation. Using the concept of receptive fields, a model of sampling representation retinal coding is developed. The coding model implements a number of known neural mechanisms, including linear/nonlinear transformations, central/surround inhibition, etc. Decoding issues are considered in the context of reconstructing spatial contrasts of images, modeling the responses of simple cells of the primary visual cortex. It is shown that the coupled ON-OFF coding model allows reconstructing sharp image details in the form of local edges. At the end of the article, to substantiate the adequacy of the synthesized coding procedure, the results of computational image delineation are demonstrated, and the perceptual quality obtained is discussed.