As randomized singular value decomposition (RSVD) is widely used in data compression, signal processing and image denoising, the increasing matrix scale puts forward higher requirements for the traditional computing platform. Therefore, a scheme of RSVD based on the spatial optical computation is proposed. The dimensions of a matrix are reduced by the inherent properties of the complex media, and there is no need to generate and store random Gaussian matrices. In this way, the computing overhead of RSVD can be effectively reduced. The experiment proves that the proposed scheme can achieve RSVD for a 80×80 matrix with a relative error of less than 0.1 when 220 mesh ground glass is used as a complex medium, the sampling rate is 0.2, and the dimension of macropixel block is 10×10. Compared with the traditional method, it effectively reduces the time complexity and space complexity of RSVD. Finally, the effect of the scheme is verified through image compression, which provides a basis for further research on large-scale image matrix algorithms.