Hyperspectral imaging has many applications including target detection, classification and environmental monitoring. Whereas, the spectral resolution and spatial resolution of hyperspectral image are mutually constrained due to the limited imaging conditions and imaging devices, so it is hard to directly acquire hyperspectral image with both high spatial resolution and spectral resolution. This paper aims to merge low-spatial resolution hyperspectral image with high-spatial resolution multispectral image to obtain high-spatial resolution hyperspectral image. The proposed algorithm divides the hyperspectral image into overlapping bands and non-overlapping bands depending on the band range of hyperspectral image and multispectral image. On the overlapping bands, a non-negative dictionary learning method is used to achieve the fusion result of hyperspectral image and multispectral image. On the non-overlapping bands, a neural network is learned to express nonlinear mapping between overlapping bands and non-overlapping bands of hyperspectral image, which is exploited to acquire the high quality non-overlapping hyperspectral bands. Finally, the processing results of overlapping and non-overlapping bands are combined to acquire a hyperspectral image with high spatial resolution. Experiments on both simulated and real datasets show that the proposed method can acquire better or comparable fusion result with the current state-of-the-art methods.