The traditional two-dimensional JND (Just Noticeable Difference) model can only estimate the minimum noticeable difference of the planar image, and is not completely suitable for the large field of view image under virtual stereo vision. First, according to the visual characteristics of the human eye, corresponding binocular observation experiments are designed for the four masking characteristics of brightness, contrast, fovea and stereo depth, and the mathematical model of JND is established through experimental data, and compared with other current JND models. The results show that this model can remove more visual redundancy under the same perceptual quality. Then the visual perception redundancy model is applied to image compression. For this purpose, a multi-level compression algorithm is proposed. The algorithm performs different levels of color level quantization on different regions of the image according to the human eye color difference threshold. The quantization process which combines the Floyd-Steinberg error dithering algorithm can remove visual redundant data. Finally, the algorithm verification is completed on the FPGA (Field-Programmable Gate Array) hardware platform. The results show that the average bit compression rate of the algorithm can reach 61.65%, which can effectively reduce the amount of transmission data required for VR (Virtual Reality) images.