Chemical exchange saturation transfer magnetic resonance imaging (CEST-MRI) is a new approach to detect tissue metabolic changes. However, constrained by its low signal-to-noise ratio and long scan time, CEST-MRI is usually only able to get low-resolution (LR) images in clinical examination, which is disadvantageous for the diagnosis of small lesions. But simple interpolation to improve the CEST-MRI resolution will cause a severe decrease in the accuracy of CEST-MRI metabolic analysis. Many deep-learning methods are also not satisfying for CEST-MRI super-resolution reconstruction (SRR). In this work, we formulate the CEST-MRI SRR as a hyperspectral image recovery problem. We adopt the Multispectral/hyperspectral-fusion (MS/HS fusion) method to solve this problem. And the optimization process is unrolled to a deep neural network. This method is named as OUSC-Net. We also propose a novel CEST-based Normalization loss (CBN loss) to improve the quantitative accuracy of the high-resolution (HR) CEST-MRI image series. The proposed method is evaluated on an ischemia rat dataset and a human dataset. The results show that the proposed method reconstructs high-quality HR CEST-MRI series even at high down-sampling rate and outperforms all other comparative SRR methods in metabolic quantification.