A bio-inspired model of image representation based on non-classical receptive fields

Biological vision systems involve complex neural layers that can represent and process information. Moreover, they are typically far more efficient than human-made machine vision systems. To obtain a non-task-dependent image representation schema, it may be valuable to simulate the early phase of the biological neural vision mechanism. We designed a neural model to simulate the non-classical receptive field of the ganglion cell and its local feedback control circuit. We found that, beyond the pixel level, our model can represent images self-adaptively and regularly. Our experimental results revealed this method was able to represent images faithfully and with a low cost. In addition, it produced compact and abstract approximations of images, and facilitated subsequent image segmentation, figure-ground separation, feature detection, and integration. This representation schema performed well for extracting spatial relationships from different components of an image, so can be applied to formalize image semantics. This system can be applied to object recognition or image classification tasks in future.