Lightweight detection of microplastic foreign bodies in sun-dried green tea: An improved YOLOv8 neural network model based on deep learning

Microplastic contamination poses a growing threat to human health through bioaccumulation and biomagnification along the tea food chain. To address this critical issue, this study focuses on the precise identification of microplastic contaminants during the processing of sun-dried green tea, aiming to safeguard food quality and consumer safety. We propose an enhanced YOLOv8-based deep learning model specifically designed to overcome the limitations of the original YOLOv8 architecture in detection accuracy, model complexity, and feature extraction efficiency. The proposed model incorporates the Inner-WIoU optimization algorithm to improve bounding box regression precision. Furthermore, we replace key components in the backbone and neck networks with Dual Convolutional (DualConv) blocks, while integrating Mixed Local Channel Attention (MLCA) and ADown modules to refine the network architecture and enable deep fusion of local and global feature representations. Extensive experimental validation demonstrates that the improved YOLOv8 model achieves outstanding performance in external testing for microplastic detection in sun-dried green tea, with Precision, Recall, mAP, and F1 scores reaching 99.12 %, 98.61 %, 99.18 %, and 98.95 %, respectively—significantly outperforming all eleven state-of-the-art object detection models evaluated in this study. Our work not only validates the efficacy and superiority of the modified YOLOv8 framework for this critical food safety task but also provides a comprehensive and efficient solution for microplastic monitoring in raw sun-dried green tea. These advancements lay a solid foundation for the modernization and intelligent transformation of the tea industry, as well as for the automation of food safety surveillance and the standardization of quality control protocols. • YOLOv8-IDMA for detecting microplastic contaminants in tea production. • Introduces Inner-WIoU to enhance small object localization accuracy. • Designs DualConv-ADown for edge preservation and noise suppression. • Embeds MLCA to boost spatial-channel feature discrimination ability.