Work

Research on a Strawberry Fruit Instance Segmentation Method Integrating Multi-Scale Features

Year: 2026

Type: article

Abstract: Small fruits and partially occluded contours make strawberry instance segmentation difficult in greenhouse images. This study develops PMH-YOLO11-seg from YOLO11l-seg to examine whether finer spatial features and auxiliary localization supervision improve mask prediction. A stride-4 P2 branch extends the original P3/P4/P5 feature hierarchy. MultiSEAM modules process P2/P3/P4, while P5 connects directly to the segmentation head. The enhanced P2 feature supplies both mask prototypes and a center heatmap branch. During training, the center branch learns box-derived Gaussian targets using error-weighted binary cross-entropy; it is bypassed during inference. Evaluation used a separate test set of 200 images containing 1132 strawberry instances. In repeated training experiments, the model with a center-loss weight of 0.4 achieved mask mAP50 and mAP50–95 of 97.41±0.07% and 84.35±0.11%, respectively (mean ± sample standard deviation). The corresponding improvements over the baseline means were 1.33 and 3.71 percentage points. In the single-run ablation, the full model exceeded P2 + MultiSEAM by only 0.03 percentage points in mAP50–95, calculated before rounding. Repeated component comparisons are needed to determine whether this small additional gain is consistent across training runs.

Source: Agriculture

DOI: 10.3390/agriculture16212402

Authors: Jingfa Yao, Zhaofei Wang, Zixuan Zhou, Yingchun Yuan

Institutions: Hebei University, Hebei Software Institute, Hebei Agricultural University

Language: English

Topic: Smart Agriculture and AI

Subfield: Plant Science

Field: Agricultural and Biological Sciences

Domain: Life Sciences

SDG: Zero hunger

Open Access status: gold

Estimated APC paid: $3,137