The aim of this paper is to construct a bio-inspired hierarchical neural network that could accurately represent visual images and facilitate follow-up processing. Our computational model adopted a ganglion cell (GC) mechanism with a receptive field that dynamically self-adjusts according to the characteristics of an input image. For each GC, a micro neural circuit and a reverse control circuit were developed to self-adaptively resize the receptive field. An array was also designed to imitate the layer of GCs that perform image representation. Results revealed that this GC array could represent images from the external environment with a low processing cost, and this nonclassical receptive field mechanism could substantially improve both segmentation and integration processing. This model enables automatic extraction of blocks from images, which makes multiscale representation feasible. Importantly, once an original pixel-level image was reorganized into a GC array, semantic-level features emerged. Because GCs, like symbols, are discrete and separable, this GC-grained compact representation is open to operations that can manipulate images partially and selectively. Thus, the GC-array model provides a basic infrastructure and allows for high-level image processing.