Authors

Abstract

In corn processing, threshed kernels often accumulate on dimly lit conveyor belts, leading to challenges in identifying damaged kernels. This study designed a sampling device integrating top and backlighting to explore optimal detection conditions. By comparing four datasets (top-lighting-dispersed, top-lighting-aggregated, backlighting-dispersed, and backlighting-aggregated), the highest accuracy (79.6%) was achieved under backlighting-dispersed conditions, validating its practicality. Furthermore, a lightweight optimization strategy was proposed for the YOLOv8 algorithm: introducing Ghost convolution to reduce computational redundancy, integrating attention mechanisms to enhance feature extraction of damaged regions, and replacing standard convolutions with depth wise separable convolutions in the backbone network. The optimized YOLOv8n_gcd model reduced floating-point operations (FLOPs) by 24%, improved inference speed (FPS) by 165%, and demonstrated enhanced robustness in densely stacked kernel scenarios. This research provides an efficient, low-cost, and adaptive solution for intelligent agricultural quality inspection, with both theoretical significance and practical engineering potential.

Abstract in Chinese

在玉米加工过程中，脱粒后籽粒常堆积于昏暗输送带，导致破损籽粒识别困难。本研究设计了一种集成顶光与背光的采样装置，通过对比顶光离散、顶光聚集、背光离散及背光聚集四类数据集，确定背光离散条件下的检测准确率最优（91.7%），验证了背光照明在实际场景的适用性。进一步针对YOLOv8算法提出轻量化改进方案：在主干网络中引入Ghost卷积以降低冗余计算，融合注意力机制强化损伤区域特征提取，并以深度可分离卷积替代标准卷积。优化后的YOLOv8n_gcd模型浮点运算量（FLOPs）减少24%，推理速度（FPS）提升160%，且在籽粒密集堆积场景下表现出更强的鲁棒性。该研究为农业智能化质检提供了高效、低耗且适应性强的解决方案，兼具理论价值与工程应用潜力。