Authors

Abstract

Wheat grains are a common type of cereal variety, and due to their large quantity and high demand, traditional manual quality inspection requires a significant amount of labor with potentially inadequate results. To address this issue, this study focuses on intact, damaged, moldy, and shriveled wheat grains, and establishes a YOLO-wheat automatic wheat grain appearance quality detection model. First, a large number of wheat grain sample images were collected, preprocessed, and annotated. Next, YOLOv5n, YOLOv8n, and YOLOv10n wheat grain object detection models were established, and the optimal model YOLOv8n was selected as the base model for automatic wheat grain appearance quality detection. To further improve wheat grain detection performance, the Dilation-wise Residual (DWR) module was integrated into the YOLOv8n network structure to enhance feature extraction from the expandable receptive field in the higher layers of the network. Additionally, the TripletAttention attention mechanism was introduced, and this improved network was named YOLO-wheat. Experimental results showed that YOLO-wheat achieved an mAP value of 91.3% in wheat grain appearance quality detection, representing a 4.3% improvement compared to the previous version. This study provides technical support for automatic wheat quality detection.

Abstract in English

小麦籽粒是一种常见的谷物类品种，且由于其数量多需求量大，在传统人工品质检测时耗费大量精力而效果不见得足够好。为解决上述问题，本研究以小麦完善粒、破损粒、发霉粒和干瘪粒为研究对象，建立YOLO-wheat小麦籽粒外观品质自动检测模型。首先，采集了大量小麦籽粒样本图像，并进行了数据预处理和标注整理。其次，建立了YOLOv5n、YOLOv8n、YOLOv10n小麦籽粒目标检测模型，从中选取最优模型YOLOv8n为小麦籽粒外观品质自动检测基础模型。为了进一步提升小麦籽粒检测性能，在YOLOv8n网络结构中使用Dilation-wise Residua(DWR)模块加强从网络高层的可扩展感受野中提取特征，并引入了TripletAttention注意力机制，将此网络命名为YOLO-wheat。实验结果表明YOLO-wheat在小麦籽粒外观品质检测中mAP值为91.3%,较改进前提升4.3% 。该研究可为小麦品质自动检测提供技术支持。