Deep Learning-Based Reconstruction of Specific Spectral Band Images Using RGB Images

Hyperspectral imaging technology has recently become an essential technology in the agricultural field, and plays an important role in analyzing the growth status, disease, stress, and quality of crops. Therefore, this technology has the potential to detect subtle changes in crops through spectral data, thereby realizing precision agriculture. However, hyperspectral equipment is expensive, has limitations in the use environment, and requires a long time to collect, process, and analyze a large amount of data. As a result, the current hyperspectral imaging system has limitations in implementing a simple and efficient imaging system. To overcome these challenges, this study designed an algorithm to reconstruct images of specific spectral bands from RGB images. The reconstruction algorithm utilized a deep learning-based image reconstruction model structure. Specifically, in order to improve the reconstruction performance, spatial and channel features were separated, and learning using a parallel structure was considered. Furthermore, The Gaussian Mixture Model was used to reinforce the main areas to be emphasized during learning. The loss function emphasizes the size of the prediction error to improve the stability and performance of the results, and a composite loss function that considers the response sensitive to small errors was used. Consequently, this study confirmed that the reconstruction algorithm can have significant reconstruction performance depending on the characteristics of a specific wavelength band or crop. Deep learning-based specific spectral band image reconstruction technology using RGB images suggests the possibility of effectively providing growth information in actual agricultural fields.