Abstract
The variety purity of crop seeds is the main quality indicator of seeds, which affects the yield and quality of crops. To achieve fast identification of maize seed varieties, this study collected images of 10 types of maize seeds, totaling 3,249 seeds. This research proposed a lightweight and small-object detection model for maize seed variety identification based on an improved YOLOv8 model: E-YOLOv8. Firstly, the backbone was replaced with FasterNet, which reduced redundant computation and memory access, allowing more efficient extraction of spatial features. Secondly, the Content-Aware ReAssembly of FEatures (CARAFE) was introduced, offering a larger receptive field and adaptive convolution kernels, which better aggregated contextual information, prevented feature loss, and improved the quality of upsampling and the accuracy of dense prediction tasks. Additionally, the Detect module was replaced with the improved Detect_EMA module, which efficiently retained information in each channel, reduced computational load, and more specifically optimized detection results. Lastly, the loss function was replaced with Inner_SIoU, which was more suitable for small-object detection tasks. Ablation experiments verified the performance of the model, and comparisons were made with YOLOv5, YOLOv6, YOLOv8, YOLOv10, and YOLOv11. The proposed E-YOLOv8 achieved a mean Average Precision (mAP) of 96.2%, a 4.4% improvement over YOLOv8, with enhancements in all other evaluation metrics. The improved E-YOLOv8 achieves an optimal balance between accuracy, speed, and resource efficiency. It features fast detection capabilities and can operate efficiently under limited storage conditions, meeting the real-time and efficiency requirements of agricultural applications. This study provided a theoretical foundation for the efficient detection of maize varieties and offered strong technical support for the intelligent and automated development of agriculture.