Authors

Abstract

To enhance real-time detection of corn breakage rate under dim conditions, this study designed a dual-light (top/backlight) sampling system. By comparing four datasets (top-scattered, top-clustered, backlight-scattered, backlight-clustered), the algorithm optimized with backlight-scattered data achieved optimal accuracy (79.6%). A lightweight YOLOv8n_gcd model was proposed, integrating Ghost convolution in the backbone to reduce redundancy, attention mechanisms for feature enhancement, and depthwise separable convolutions in the neck. The optimized model reduced FLOPs by 24% and increased FPS by 165%, offering an efficient, low-cost solution for agricultural quality inspection with theoretical and practical value

Abstract in Chinese

为了增强在昏暗条件下玉米破损率的实时检测，本研究设计了一种双光（顶部/背光）采样系统。通过比较四个数据集（顶光-籽粒分散、顶光-籽粒聚集、背光-籽粒分散、背光籽粒聚集），得出背光籽粒分散的数据优化的算法达到了最佳（79.6%）之后利用该数据集训练出一种轻量级的YOLOv8n_gcd模型，将Ghost卷积集成在骨干网中以减少冗余，将注意力机制用于特征增强，并在颈部进行深度可分离卷积。优化后的模型将FLOP降低了24%，FPS提高了165%，为农业质量检测提供了一种高效、低成本的解决方案，具有理论和实践价值