An object-based framework for identifying moldy corn using hyperspectral images.

Hyperspectral imaging (HSI), leveraging its non-destructive and rapid detection capabilities, has become a pivotal technology for mold detection of corn and other grains. Currently, prevalent methodologies for grain mold detection are broadly classified into pixel-based (PB) and kernel-based (KB) methods, but both have inherent limitations. PB methods depend only on spectral data for mold identification, ignoring spatial features and causing misclassifications of substances with similar spectra. KB methods extract overall kernel features to detect mold but overlook fine-grained corn characteristics, making it hard to capture mold features when mold pixels are scarce. To address these issues and improve the accuracy of moldy corn detection, this study proposes an innovative object-based (OB) framework for identifying moldy corn. In this framework, a Normalized Corn Mold Index (NCMI) for near-infrared (NIR) hyperspectral images is proposed to enhance the separability of healthy and moldy ‌corn kernels. Subsequently, multiresolution segmentation partitions each corn kernel into homogeneous objects, forming a structured foundation for feature extraction. A multi-scale convolutional network (MSCNN) is then designed to extract hierarchical features from these objects, effectively capturing fine-grained details and contextual information across scales. The extracted features are fed into a support vector machine (SVM) for final classification, constructing an MSCNN-SVM model that integrates multi-scale feature extraction with discriminative classification. Experiments show our OB framework achieves 97.01% accuracy, outperforming PB and KB methods. The newly developed framework offers robust technical support for the intelligent monitoring of food quality and safety, demonstrating extensive application prospects and significant promotion value.