Hyperspectral imaging is a powerful method for examining the chemical makeup of food and agricultural products. Nonetheless, it is expensive and complex, which restricts its practical use. Researchers from the University of Illinois Urbana-Champaign have created a technique to generate hyperspectral images from standard RGB images through advanced deep learning methods.
Hyperspectral imaging is an effective means of studying the chemical components of food and agricultural items. Unfortunately, its high costs and complexity limit practical implementation. A group of researchers at the University of Illinois Urbana-Champaign has devised a technique to reconstruct hyperspectral images from ordinary RGB images using deep machine learning. This innovation could significantly simplify the analytical process and potentially transform product evaluation in agriculture.
“Hyperspectral imaging requires expensive equipment. If we can utilize RGB images taken with a regular camera or smartphone, we can predict product quality using a low-cost, handheld device,” said Md Toukir Ahmed, the lead author and a doctoral student in the Department of Agricultural and Biological Engineering (ABE) at the College of Agricultural, Consumer and Environmental Sciences and The Grainger College of Engineering at Illinois.
The researchers evaluated their technique by examining the chemical composition of sweet potatoes. They studied soluble solid content in one instance and dry matter in another—both critical factors affecting the flavor, nutritional value, desirability in the market, and suitability for processing sweet potatoes. By employing deep learning models, they transformed the data from RGB images into hyperspectral images.
“RGB images only show visible features such as color, shape, size, and surface imperfections; they do not reveal any chemical characteristics. RGB images capture wavelengths ranging from 400 to 700 nanometers and comprise three channels—red, green, and blue. In contrast, hyperspectral images encompass many channels and wavelengths from 700 to 1000 nm. Using deep learning techniques, we can map and recreate this extended range, allowing us to identify chemical attributes from RGB images,” explained Mohammed Kamruzzaman, an assistant professor in ABE and co-author of both papers.
Hyperspectral imaging records intricate spectral signatures at various spatial points across hundreds of narrow bands, which combine to produce hypercubes. Utilizing innovative deep learning algorithms, Kamruzzaman and Ahmed were able to devise a model that reconstructs hypercubes from RGB images to extract relevant product analysis information.
They calibrated their spectral model with reconstructed hyperspectral images of sweet potatoes, achieving an accuracy of over 70% in predicting soluble solid content and 88% in determining dry matter content, which represents a significant advancement over earlier studies.
In a third publication, the research group employed deep learning strategies to recreate hyperspectral images that forecast chick embryo mortality, relevant to the egg and hatchery sector. They investigated various methods and provided recommendations for the most precise approaches.
“Our findings hold significant promise for transforming the assessment of agricultural product quality. By generating detailed chemical information from straightforward RGB images, we are opening up new avenues for cost-effective and accessible analysis. Although there are hurdles to overcome in scaling this technology for industrial application, its potential to reshape quality control in agriculture is an incredibly promising venture,” Kamruzzaman concluded.
