基于改进YOLO v4的玉米种子外观品质检测方法

针对玉米种子在外观品质检测中需要快速识别与定位的需求,提出了一种基于改进YOLO v4的目标检测模型,同时结合四通道(RGB+NIR)多光谱图像,对玉米种子外观品质进行了识别与分类。为了减少改进后模型的参数量,本文将主干特征提取网络替换为轻量级网络MobileNet V1。为了进一步提升模型的性能,通过试验研究了空间金字塔池化(Spatial pyramid pooling, SPP)结构在不同位置上对模型性能的影响,最终选取改进YOLO v4-MobileNet V1模型对玉米种子外观品质进行检测。试验结果表明,模型的综合评价指标平均F1值和mAP达到93.09%和98.02%,平均每检测1幅图像耗时1.85 s,平均每检测1粒玉米种子耗时0.088 s,模型参数量压缩为原始模型的20%。四通道多光谱图像的光谱波段可扩展到可见光范围之外,并能够提取出更具有代表性的特征信息,并且改进后的模型具有鲁棒性强、实时性好、轻量化的优点,为实现种子的高通量质量检测和优选分级提供了参考。

Corn Seed Appearance Quality Estimation Based on Improved YOLO v4

Aim to identify and position corn seed, an object detection model based on improved YOLO v4 was proposed. This model combined with multi-spectral images with four channels (RGB+NIR), the appearance quality of corn seeds was identified and classified. In order to reduce the number of parameters in the model, the trunk feature extraction network was replaced with the lightweight network MobileNet V1. To improve the performance of the model, the effect of spatial pyramid pooling (SPP) structure on the model performance was studied. Finally, the improved YOLO v4-MobileNet V1 model was selected to detect the appearance quality of corn seeds. The experimental results showed that the comprehensive evaluation indexes F1 and mAP of the model reached 93.09% and 98.02%, respectively. The average detection time of each image was 1.85s, and the average detection time of each corn seed was 0.088s. And the number of model parameters was compressed to 20% of the original model. The spectral band of four channel multi-spectral image can be extended beyond the visible range. Image can extract more representative feature information. The improved model had the advantages of strong robustness, good real-time performance and lightweight. It can provide a reference for high throughput quality detection and optimal classification of seeds.