With the rapid advancement of spectral imaging technology, the use of multispectral filter array (MSFA) to collect the spatial and spectral information of multispectral images has become a research hotspot. The uses of the original data are limited because of its low sampling rate and strong spectral inter-correlation for reconstruction. Therefore, we propose a multi-branch attention residual network model for spatial-spectral association based on an 8-band 4 × 4 MSFA with all-pass bands. First, the multi-branch model was used to learn the image features after interpolation in each band; second, the feature information of the eight bands and the all-pass band were united by the spatial channel attention model designed in this paper, and the application of multi-layer convolution and the convolutional attention module and the use of residual compensation effectively compensated the color difference of each band and enriched the edge texture-related feature information. Finally, the preliminary interpolated full-pass band and the rest of the band feature information were used for feature learning of the spatial and spectral correlations of multispectral images through residual dense blocks without batch normalization to match the spectral information of each band. Experimental results show that the peak signal-to-noise ratio, structural similarity, and spectral angular similarity of the test image under the D65 light source outperform the state-of-the-art deep learning method by 3.46%, 0.27%, and 6%, respectively. This method not only reduces artifacts but also obtains more texture details.