Abstract: Dual-energy X-ray imaging technology, which utilizes the attenuation differences of high-energy and low-energy X-rays in different materials, can achieve the identification of the inspected objects and is widely used in the field of large-scale cargo/vehicle security inspection. However, traditional dual-energy systems operating at 3-9 MeV suffer from insufficient sensitivity to thin materials below 20 g/cm². These limitations arise because conventional integrating detectors simultaneously collect photons across both the MeV range (where Compton scattering dominates and attenuation is nearly material-independent) and the keV range (where the photoelectric effect dominates and attenuation is highly atomic-number-sensitive), causing mutual interference between the two energy regimes. A multi-energy imaging technique was proposed based on a 6, 3, and 0.3 MeV tri-energy accelerator and a dual-layer detector, which effectively overcomes the constraints on material classification imposed by spectral crosstalk, expanding the range of classifiable materials while improving classification accuracy. The material classification experiments were conducted on polyethylene, aluminum, iron, and lead-antimony alloy blocks spanning 1-120 g/cm². Four supervised machine learning algorithms were trained and compared using Z-score normalized features. The random forest classifier reaches 92% overall recall across 1-120 g/cm², a 28% improvement over traditional dual-energy systems. Feature importance analysis confirms that the 0.3 MeV first-layer channel contributes most to classification performance owing to its dominant photoelectric interactions.