An image representation of infrastructure based on non-classical receptive field

Biological vision systems have become highly optimized over millions of years of evolution, developing complex neural structures to represent and process stimuli. Moreover, biological systems of vision are typically far more efficient than current human-made machine vision systems. The present report describes a non-task-dependent image representation schema that simulates the early phase of a biological neural vision mechanism. We designed a neural model involving multiple types of computational units to simulate ganglion cells and their non-classical receptive fields, local feedback control circuits and receptive field dynamic self-adjustment mechanisms in the retina. We found that, beyond the pixel level, our model was able to represent images self-adaptively and rapidly. A series of statistical analyses revealed that this model not only produces compact and abstract approximations of images, but also retains their primary visual features. In addition, the improved representation was found to substantially facilitate contour detection and image segmentation. We propose that this improvement arose because ganglion cells can resize their receptive fields, enabling multi-scale analysis functionality, a neighborhood referring function and a localized synthesis function. The ganglion cell layer is the starting point of subsequent diverse visual processing. The universality of this cell type and its functional mechanisms suggests that it will be useful for designing image processing algorithms in future.