One of the imaging techniques to produce spectral information at the Near Infrared spectrum range is hyperspectral imaging. To minimize the high cost and complicated imaging techniques, hyperspectral image reconstruction is performed from RGB images. The HR-ResNet algorithm uses residual blocks by utilizing shortcut connections to reduce the vanishing of gradients and produce optimal model performance. Using the right resblock layer and the Batch Normalization layer can also speed up training time thereby increasing the performance of the reconstruction model. The performance evaluation will be tested using 2 evaluation metrics RMSE and MAE. Image acquisition was performed using a hyperspectral camera with spectral range of 400-1000 nm. The RGB image used as input was obtained by converting the image using the CIE 1931 color matching function. The dataset comparison between training, validating, and testing used was 50:25:25. Variation of the band numbers of target and image spatial size was also carried out to determine the performance of the reconstruction model. Based on the results, it can be seen that the reconstruction model is able to reconstruct hyperspectral images from RGB images with RMSE and MAE errors of 1.20 and 0.61, respectively. The variation in the band numbers of target also affects the performance of the model because the reconstruction can work better if using a smaller number of reconstruction target bands, while variations in image spatial size do not significantly affect the performance of the reconstruction model.