The invention discloses a non-linear compression spectral imaging method based on hidden space worked example learning, and mainly solves the problems of over high time complexity and space complexity in dictionary learning under a non-linear space by using a kernel function in the prior art. The implementing steps include a first step of pre-processing a training sample, and obtaining a virtual training sample; a second step of training the virtual training sample via a linear dictionary learning method, and obtaining a sparse dictionary; a third step of initializing an observation matrix randomly, and achieving non-linear compression sensing spectral imaging through a kernel compression sensing method; and a fourth step of recovering an original signal by using a pre-image method. The experimental result shows that the method of the invention has better reconstruction results and much lower time complexity at the same sampling rate as compared with a conventional KPCA dictionary learning method, and can be used for low rate sampling and recovery of hyperspectral images.