Lightweight MSW-YOLOv8n-Seg: the instance segmentation of maturity on cherry tomato with improved YOLOv8n-Seg

TL;DR

This paper introduces a lightweight model for segmenting cherry tomato maturity levels in natural environments, improving accuracy and speed for automated picking.

Contribution

A novel lightweight instance segmentation model, MSW-YOLOv8n-Seg, is proposed for cherry tomato maturity detection with improved accuracy and real-time performance.

Findings

The model achieved 90.8% precision, 86.3% recall, and 83.9% mAP@0.5 on test sets.

MSW-YOLOv8n-Seg outperformed existing models in precision, recall, and mAP@0.5 by significant margins.

The model has a high inference speed of 52.9 FPS and low latency of 18.2ms, suitable for real-time applications.

Abstract

Automatic and accurate segmentation of cherry tomato maturity in natural environment is the foundation for automatic picking. Lacking of significant differences in adjacent maturity and the problem of mutual occlusion between fruits usually affect the picking process. According to the changes in phenotypic characteristics of cherry tomato during its mature period and the Chinese national standard GH/T 1193-2021, a lightweight maturity instance segmentation method of cherry tomato with 5 levels, including green, turning, pink, light red and red was proposed based on improved YOLOv8n-Seg model, named as MobileViTv3-SK-WIoU-YOLOv8n-Seg (MSW-YOLOv8n-Seg).

In this model, MobileViTv3 was introduced into the original YOLOv8 model as backbone for feature extraction to reduce the parameters of the original model; selective kernel (SK) attention module was added to the neck part to improve the feature expression ability of the model; the complete intersection over union (CIoU) loss function in the original head part was replaced with wise intersection over union (WIoU), which can effectively filter low-quality samples and improve the stability and reliability of the model in complex scenes. The proposed model can better balance the relationship between segmentation speed, accuracy, and model computational complexity.

The experimental results show that the bounding box precision, recall and mean average precision (mAP)@0.5 of the improved model on the test sets were 90.8%, 86.3% and 83.9% respectively, and the model size was 6.0 MB. Compared with YOLOv7-Mask, YOLOv8n-Seg, YOLOv9s-Seg, YOLO11n-Seg, Mask R-CNN (Mask region-based convolutional neural network) and Mask2Former, the bounding box precision increased by 9.6%, 5.2%, 5.7%, 12.3%, 13.3% and 5.0%, the recall increased by 7.8%, 7.4%, 8.8%, 13.1%, 13.9% and 0.1%, and the mAP@0.5 increased by 10.5%, 3.0%, 0.9%, 15.0%, 13.8% and 1.4% respectively. In terms of inference speed, the MSW-YOLOv8n-Seg has the highest inference speed, with FPS of up to 52.9 f·s-1 and latency of only 18.2ms, which demonstrates its real-time processing capability.

The results show that the improved MSW-YOLOv8n-Seg model is optimal, and it suitable for instance segmentation scenarios with high real-time performance and can provide effective exploration for automated cherry tomato fruit picking.