The present invention discloses a multi-layer gradient constraint regression-based single-image super-resolution reconstruction method. The method mainly comprises the following steps of by using a training image, respectively learning a multi-layer regression model in the gradient and residual domain; inputting a low-resolution image interpolation to obtain an initial high-resolution image; adopting the current high-resolution image as the input, and respectively estimating the gradient and the residual error based on a gradient and residual domain regression model corresponding to the current layer; according to the estimated gradient and the estimated residual error of the current layer, constructing and optimizing a reconstruction cost function and obtaining the reconstruction result of the current layer; adopting the reconstruction result of a previous layer as a new high-resolution image, and repeating the above two steps until reaching a maximum number of reconstruction; and finally outputting a final reconstruction result. According to the technical scheme of the invention, a high-quality high-resolution image can be reconstructed based on a low-resolution image, and the reconstruction speed is faster. The method can be applied to the fields of military, remote sensing, medical treatment and the like.