IMPROVED YOLOV8-ALGORITHM FOR SORTING FRESH WHITE TEA: COMBINING FEATURE ENHANCEMENT AND ATTENTION

改进的 YOLOV8 新鲜白茶分选算法：将特征增强与注意力相结合

DOI : https://doi.org/10.35633/inmateh-75-80

Authors

Xuedong YU

1) College of Intelligent Manufacturing and Elevator Technology, Huzhou Vocational and Technical College;2)Huzhou Key Laboratory of Robot System Integration and Intelligent Equipment

Yadong NIU

College of Intelligent Manufacturing and Elevator Technology, Huzhou Vocational and Technical College

Zhongyou ZHOU

1) College of Intelligent Manufacturing and Elevator Technology, Huzhou Vocational and Technical College;2)Huzhou Key Laboratory of Robot System Integration and Intelligent Equipment

Bo LU

1) College of Intelligent Manufacturing and Elevator Technology, Huzhou Vocational and Technical College;2)Huzhou Key Laboratory of Robot System Integration and Intelligent Equipment

Kaiqiang JIN

1) College of Intelligent Manufacturing and Elevator Technology, Huzhou Vocational and Technical College; 2)Huzhou Key Laboratory of Robot System Integration and Intelligent Equipment

(*) Rongyang WANG

1) College of Intelligent Manufacturing and Elevator Technology, Huzhou Vocational and Technical College; 2)Huzhou Key Laboratory of Robot System Integration and Intelligent Equipment

(*) Corresponding authors:

rongyang1987@126.com |

Rongyang WANG

Abstract

In this paper, an improved intelligent sorting algorithm for YOLOv8 white tea fresh leaves is proposed to solve the problems of unclear tea grades and uneven product levels caused by mechanical picking. The algorithm introduces the Dynamic Snake Convolution module (DSConv) for feature enhancement and adds an attention mechanism module, the Multi-Head Self-Attention (MHSA). Experiments show that the YOLOv8-DsConv-MHSA algorithm has an average accuracy of 96.4% and an average detection rate of 126.6 FPS per second, which is the best algorithm for white tea fresh leaf sorting in the comprehensive comparison.

Abstract in Chinese

本文提出了一种改进的 YOLOv8 白茶鲜叶智能分拣算法，以解决机械采摘造成的茶叶等级划分不清和产品层次不齐的问题。该算法引入了动态蛇形卷积模块（DSConv）进行特征增强，并且添加了注意力机制模块--多头自我注意力（MHSA）。实验表明：YOLOv8-DsConv-MHSA算法的平均准确率为96.4%，平均检测率为每秒126.6 FPS，是综合比较中白茶鲜叶分选效果最好的算法。

Indexed in