Authors

Abstract

This study proposes a grading method using combined target detection and segmentation models to enhance grading accuracy for pepper plug seedlings. We constructed separate datasets for target detection and image segmentation of pepper plug seedlings, then trained various detection models including EfficientDet, Faster R-CNN, SSD, and YOLO-series architectures. After comprehensive evaluation, YOLOv5 achieved optimal performance with 99.5% mAP, a compact 3.8 MB model size, and rapid 5.53 ms inference time per image. To avoid the impact of the culture medium and adjacent pepper seedlings on the image, we developed an improved U-Net model incorporating an Efficient Channel Attention (ECA) mechanism, enhancing segmentation accuracy by 1.29% to 93.01% while reducing processing time by 69.7% (33.32 ms). Subsequent feature extraction analyzed lateral area, height, stem thickness, hypocotyl length, and divergence degree from segmented images. Using these extracted features, grading models were trained employing support vector machines, k-nearest neighbors, and random forests. The random forest model achieved a grading accuracy of 99.33%, validating that this method meets the accuracy requirements for pepper seedling grading.

Abstract in Chinese

为了提高辣椒穴盘苗分级检测精度，本研究提出了一种基于目标检测和分割模型相结合的辣椒穴盘苗分级方法。首先，构建辣椒穴盘苗目标检测和图像分割两个数据集；其次，训练了Efficientdet，Faster-Rcnn，SSD以及YOLO系列目标检测模型，经过综合对比YOLOv5检测效果最好，平均精度均值（mean average precision，mAP）达到99.5%，模型大小3.8 MB，单张图像检测时间5.53ms；然后，为了避免培养基质以及相邻辣椒苗对图像的影响，使用U-Net算法作为本文的图像处理算法的基础模型，并引入通道注意力机制ECA（Efficient Channel Attention Network）进行改进，与U-Net相比改进后模型分割准确率提高1.29%，达到93.01%，检测时间缩短69.7%，为33.32ms；最后基于图像分割算法处理后的图像，提取了辣椒穴盘苗面积，高度，茎粗，下胚轴长度和发散程度等参数，并利用提取的特征参数完成了支持向量机，K最邻近算法，和随机森林算法的分级模型训练，其中随机森林模型分级准确率为99.33%，验证了该方法能够满足辣椒穴盘苗分级精度要求。