Authors

Abstract

Pod pepper(Capsicum annuum var. conoides), a common variety of chili pepper, poses a challenge for traditional object detection methods due to its complex morphological features and diverse types. This study focuses on the application of machine vision technology to address the issue of pod pepper object detection. Firstly, a large number of pod pepper sample images were collected, followed by data preprocessing and annotation. Subsequently, YOLOv3, YOLOv5, YOLOv6, and YOLOv8 pod pepper object detection models were established, with YOLOv8 yielding the best detection results with a mean Average Precision (mAP) value of 81.6%. Next, different attention mechanisms were incorporated into the YOLOv8 network structure, with experimental results indicating that the Triplet Attention mechanism performed the best in pod pepper object detection, achieving an mAP value of 82.5%, a 0.9% improvement over YOLOv8. To further optimize the effectiveness of the attention mechanisms, Triplet Attention was added at different positions within the YOLOv8 network. The experiment showed that the location of adding the attention mechanism significantly impacted the pod pepper detection results. When Triplet Attention was added at the 5th layer, the best detection performance was achieved, with an mAP value of 84.1%, a 2.5% improvement over the original YOLOv8. This research provides technical support for intelligent harvesting of pod pepper.

Abstract in Chinese

朝天椒是一种常见的辣椒品种，由于其形态特征复杂且种类较多，传统的目标检测方法在其识别方面存在一定的挑战。本研究基于机器视觉技术，针对朝天椒目标检测问题展开研究。首先，采集了大量朝天椒样本图像，并进行了数据预处理和标注整理。其次，建立了YOLOv3、YOLOv5、YOLOv6、YOLOv8朝天椒目标检测模型，对比不同检测模型效果，YOLOv8检测结果最优，检测mAP值为81.6%。然后在YOLOv8网络结构中添加不同注意力机制，实验结果表明Triplet Attention机制在朝天椒目标检测中表现最好，检测结果mAP值为82.5%，比YOLOv8提升0.9%。为了进一步验证注意力机制的效果最大优化性，将Triplet Attention添加到YOLOv8网络不同位置，试验结果表明添加注意力机制的位置对朝天椒检测结果有显著影响。当Triplet Attention添加到5层，检测效果最好，检测mAP值为84.1%，相比原始YOLOv8提升2.5%。该研究可为朝天椒智能采摘提供技术支持。