December 2, 2025Open Access

MDE-DETR: multi-domain enhanced feature fusion algorithm for bayberry detection and counting in complex orchards

Key Points

MDE-DETR demonstrates improved detection accuracy for bayberries, enhancing yield predictions in orchards.
The algorithm features a novel backbone and fusion network tailored for small and occluded targets.
In tests, MDE-DETR outperformed the baseline model, significantly increasing accuracy and reducing resource usage.
This method may enhance automated fruit counting in densely planted environments, indicating its practical application.

Abstract

Introduction Bayberry detection plays a crucial role in yield prediction. However, bayberry targets in complex orchard environments present significant detection challenges, including small volume, severe occlusion, and dense distribution, making traditional methods inadequate for practical applications. Methods This study proposes a Multi-Domain Enhanced DETR (MDE-DETR) detection algorithm based on multi-domain enhanced feature fusion. First, an Enhanced Feature Extraction Network (EFENet) backbone is constructed, which incorporates Multi-Path Feature Enhancement Module (MFEM) and reparameterized convolution techniques to enhance feature perception capabilities while reducing model parameters. Second, a Multi-Domain Feature Fusion Network (MDFFN) architecture is designed, integrating SPDConv spatial pixel rearrangement, Cross-Stage Multi-Kernel Block (CMKBlock), and dual-domain attention mechanisms to achieve multi-scale feature fusion and improve small target detection performance. Third, an Adaptive Deformable Sampling (ADSample) downsampling module is constructed, which dynamically adjusts sampling positions through learnable spatial offset prediction to enhance model robustness for occluded and dense targets. Results and discussion Experimental results demonstrate that on a self-constructed bayberry dataset, MDE-DETR achieves improvements of 3.8% and 5.1% in mAP50 and mAP50:95 respectively compared to the RT-DETR baseline model, reaching detection accuracies of 92.9% and 67.9%, while reducing parameters and memory usage by 25.76% and 25.14% respectively. Generalization experiments on VisDrone2019 (a small-target dataset) and TomatoPlantfactoryDataset (a dense occlusion dataset) datasets further validate the algorithm's effectiveness, providing an efficient and lightweight solution for small-target bayberry detection in complex environments.

Ask AI

Helpful

Bookmark

View Full Paper