<p>This study developed a recognition model for imperfect maize kernels based on hyperspectral imaging technology. Imperfect maize kernel samples, including insect-damaged, spotted, broken, sprouted, moldy, and heat-damaged kernels, were prepared following established standards. A near-infrared hyperspectral imaging system was employed to acquire images under controlled parameters, with black and white reference corrections applied to reduce interference. Spectral features were extracted using ENVI software, and a multilayer perceptron recognition model containing 105 hidden layers neurons was constructed. Model stability and performance were enhanced through methods such as 5-fold cross-validation. Strict and standardized operations were carried out during sample collection, and correction techniques were used to obtain high-quality data. The multilayer perceptron neural network achieved a recognition accuracy of 96.5%, demonstrating the importance of selecting appropriate characteristic bands. While each model had their own strengths and limitations, the neural network exhibited a long training time but high accuracy. In practical applications, the model structure and analysis functions need to be optimized to improve the recognition speed. This study provides effective models and methods for the recognition of imperfect maize kernels and lays a foundation for subsequent optimization of the system.</p>