Research on Nondestructive Classification Methods for Coated Corn Varieties Based on Multi‐Scale Feature Extraction

ABSTRACT To achieve rapid and accurate classification of coated corn varieties, a dataset containing five types of coated corn was constructed, with an original resolution of 640 × 640 pixels, comprising a total of 6715 images. In order to reduce the excessive reliance of classification models on dataset preprocessing, three improved modules—namely the ES‐FCA Block structure, LSK‐FourBranch Model structure, and WP convolution structure—were proposed based on the YOLOv8 model framework. These three improved modules were sequentially combined to build a new backbone network, leading to the development of a lightweight classification model with multistage downsampling for feature extraction (YOLOv8‐ES‐LSK), which is specifically designed for the classification of coated corn varieties. Compared with deep learning network models such as EfficientNet, EfficientViT, ViT, ResNet50, and YOLOv11‐cls, the proposed model has demonstrated excellent experimental performance, with a maximum classification accuracy of 97.99%, and the number of parameters and computational complexity are 0.77 M and 1.8 FLOPs, respectively. Experimental results indicate that the approach of multiple downsampling‐based feature extractions enables effective extraction of features from coated corn varieties with high similarity. Meanwhile, the proposed model achieves both efficient classification and lightweight and efficient performance. Specifically, model inference conducted via ONNX Runtime has yielded favorable classification results, laying a solid foundation for the potential deployment of the model on edge devices. Furthermore, we have also developed a complete set of coated corn sorting system, along with its accompanying nondestructive testing (NDT) software platform.