Application of CNN-based hyperspectral image recognition technology for rapid identification of moldy grains

Abstract

To address the urgent need for rapid, non-destructive, and full-scale inspection of moldy maize in grain storage scenarios, this study develops an online grading model based on hyperspectral imaging and a lightweight convolutional neural network. Using hyperspectral data cubes in the 400–1000 nm range as input, a third-order spatial–spectral joint convolution architecture is designed to enable mold severity classification from grade 0 to 4 at an 8×8×128 feature representation level. Trained on 250 laboratory samples, the model achieves an overall accuracy of 98.5% on an independent test set, with an F1-score of 98.0% for the critical grade 3 category. The inference time per kernel is less than 50 ms. In a blind test of 100 samples conducted at the Beilin National Grain Depot, the model demonstrates a consistency rate of 96.0% with manual inspection, confirming its capability to perform real-time identification of highrisk kernels within conveyor belt processing constraints. This provides a feasible technical pathway for intercepting mycotoxin contamination prior to grain storage.