Coded aperture spectral imaging is a snapshot spectral imaging method, but it usually has the problems of large reconstruction error and high reconstruction computational complexity. To solve this problem, this paper proposes a compressed spectral reconstruction method based on non-local sparse representation and dual-camera system. First, a dual camera system is used to obtain the spectral and spatial data of the target. This dual camera system has two branches, the light is divided into two paths through a spectroscope, half enters the coded aperture spectral imaging system to obtain encoded images, and the other half is received by an RGB camera to obtain RGB images. The RGB observation image is used to construct 3D image patches, and k-means clustering is used to classify these 3D image patches. Then we propose a method to estimate the non-local similarity of target spectral image by RGB observation. The clustering and similarity estimation results of 3D image patches are used to guide the classification and similarity estimation of target spectral images. Divide the initialized target spectral image into a series of three-dimensional spectral patches, and classify the spectral patches based on the previous clustering results. Perform principal component analysis on each cluster, obtain the common features between different patches of the target spectral image, and use them to sparsely represent other spectral patches. For each patch, the sparse representation coefficients of the current patches are estimated by the weighted sum of sparse representation coefficients of nonlocal similar patches, and the weighted coefficients are calculated from the 3D image patches constructed by RGB observation. In order to improve the reconstruction quality, we set adaptive regularization parameters for sparse representation coefficients. We transform these operations into a variational optimization model, and then adopt an alternative optimization scheme to solve the objective function. We use conjugate gradient descent method and iterative threshold shrinkage method to optimize alternately. After every fifteen iterations, perform a principal component analysis on the classified three-dimensional spectral patches to obtain a new dictionary, and continue to repeat the iterative process. Through multiple repetitions, the final objective function converges, and the reconstructed spectral image can be obtained. We have done simulation experiments on the public spectral image dataset, and the experimental results show that our method has smaller spatial and spectral dimensions errors than other methods. We conducted simulation experiments on public spectral datasets, and the results show that our method has smaller errors in both spatial and spectral dimensions. From the perspective of spatial dimension, the proposed method can retain more details. From the spectral dimension, the method has smaller error and smaller error fluctuation in almost all wavebands than other methods. In addition, we compare the RGB auxiliary dictionary learning and similarity estimation method proposed in this paper with the common intermediate result dictionary learning and similarity estimation methods, the RGB auxiliary reconstruction method saves nearly half of the time while maintaining the same reconstruction quality. Finally, we set up an imaging system to do experiments on real data, and took images with filters for reference. The experiments show that our method can also obtain the best reconstruction quality on real data, which is most similar to the images obtained with filters. We also analyzed the influence of some factors, such as sampling step size and patch size, and selected the most appropriate parameter settings through a large number of experiments. Experiments on simulation data and real data show that our reconstruction model can greatly improve the reconstruction quality of spectral images in spatial and spectral dimensions, and the RGB observation assisted reconstruction method can effectively reduce the reconstruction time.