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

Area of Science:

• Computer Vision
• Agricultural Technology
• Machine Learning

Background:

• Bayberry detection is vital for accurate yield prediction.
• Traditional methods struggle with small, occluded, and densely distributed bayberry targets in complex orchard environments.

Purpose of the Study:

• To develop an advanced detection algorithm for bayberry yield prediction in challenging orchard conditions.
• To enhance the accuracy and efficiency of detecting small and occluded bayberry targets.

Main Methods:

• Proposed a Multi-Domain Enhanced DETR (MDE-DETR) algorithm featuring an Enhanced Feature Extraction Network (EFENet) with Multi-Path Feature Enhancement Module (MFEM).
• Implemented a Multi-Domain Feature Fusion Network (MDFFN) incorporating SPDConv, Cross-Stage Multi-Kernel Block (CMKBlock), and dual-domain attention for multi-scale feature fusion.
• Introduced an Adaptive Deformable Sampling (ADSample) module to improve robustness against occlusion and dense target distributions.

Main Results:

• MDE-DETR achieved 92.9% mAP50 and 67.9% mAP50:95 on a bayberry dataset, outperforming RT-DETR by 3.8% and 5.1% respectively.
• Reduced model parameters by 25.76% and memory usage by 25.14%, offering an efficient and lightweight solution.
• Demonstrated strong generalization on small-target (VisDrone2019) and dense occlusion (TomatoPlantfactoryDataset) datasets.

Conclusions:

• The MDE-DETR algorithm provides an effective and efficient solution for bayberry detection in complex agricultural settings.
• The proposed methods significantly improve the detection of small, occluded, and dense targets, crucial for precision agriculture and yield prediction.