A task-specific architecture with multi-scale attention and shape-aware loss for strawberry phenophase recognition in...

To address the challenges of recognizing small strawberry targets and achieving accurate phenological perception in complex field environments, this paper proposes a novel end-to-end lightweight detection architecture named HCMS-Net. The backbone is a Residual Efficient Layer Aggregation Network (R-ELAN) enhanced with a Multi-Scale Convolutional Attention (MSCA) mechanism, which emphasizes subtle color and texture variations to differentiate key phenological phases. For feature fusion, hypergraph convolution (from HyperC2Net) and a Mixed Aggregation Network (MANet) are incorporated, modeling the clustered morphology of strawberries and strengthening the representation of sparse small fruits. The detection head incorporates a lightweight Conv2Former module to capture long-range dependencies and spatial contextual information across growth stages, thereby enhancing the model's capacity to represent continuous phenological changes. A Shape-Normalized Wasserstein Distance (Shape-NWD) loss is introduced to stabilize optimization against minor pixel deviations. Experimental results demonstrated that HCMS-Net achieved a mean average precision (mAP) of 94.9% and an F1-score of 90.0%. Specifically, the average precision (AP) values for the flowering, young fruit, green fruit, veraison, and mature fruit stages reached 99.3%, 88.3%, 90.9%, 97.0%, and 98.2%, respectively. Heatmaps confirmed HCMS-Net's precise attention focus across all five phenological stages, effectively suppressing irrelevant backgrounds. Compared to ten mainstream detectors, HCMS-Net surpassed alternatives such as RT-DETR and the YOLOv5n to v13n by 3.4-8.0 percentage points in mAP. It even surpassed YOLOv12s by 2.7 percentage points, while containing only 32.86% of its parameters. The model offers high accuracy and efficiency for phenological period detection, supporting selective harvesting and intelligent agricultural management.