Viacheslav Antsiperov, Kotelnikov Institute of Radioengineering and Electronics, Russia

This work discusses the image neuromorphic encoding/decoding issues inspired by the mechanisms of visual perception encoding/decoding. The importance of this topic is directly related to the current problems of perceptual quality and perceptual reconstruction of images today. Therefore, to obtain reliable results in these directions, it was natural to turn to the most adequate mechanisms of perception. As a result, we propose a new approach to image processing, which uses the most realistic representation of the input data in the form of a stream of events or counts. Such events / counts simulate the firing of retinal receptors in response to the action of radiation recorded. The statistical model of the counts stream is chosen in the form of the two-dimensional inhomogeneous Poisson point processes, considered as a convenient representation of the input data. In the current paper such a representation is referred to as sampling representation. To adequately model the mechanisms of neural encoding of input data, we consistently use the concept of receptive fields. This general model implements well-known features of neural processing, including central/lateral inhibition. Decoding issues are considered in the context of Retinex paradigm of contrast detection. It is shown that the model of coupled ON-OFF receptive fields allows to restore sharp image details in the form of local edges. At the end of the work, we demonstrate an interpretation of synthesised encoding/decoding as classical smoothing and edge outlining of encoded images

Neuromorphic Systems, Sampling Representation, Neural Encoding, Receptive Fields, Edges Detection