Answers are grounded in this paper and cite their sources.
The detection of small, occluded, early-stage pomegranate fruits within complex orchard environments poses a significant challenge for efficient deployment on edge devices. To overcome this, we propose PG-DETR, a lightweight end-to-end Detection Transformer (DETR). Our model introduces two key innovations: a highly efficient lite HGNetv2 backbone that drastically reduces model complexity, and a novel multi-path hierarchical context fusion encoder (MHCF Encoder) that enhances multi-scale feature fusion to improve robustness. Experimental results on the PG dataset show that PG-DETR achieves a mean average precision at 50% IoU (mAP50) of 93.37% with only 6.7 million parameters and a computational load of 12.4 gigaflops per second (GFLOPs). Notably, compared to the RT-DETR-R18 baseline, our model reduces parameters by 66.2% and GFLOPs by 78.2%, while increasing inference speed by a factor of 3.2, reaching 173.58 frames per second (FPS). PG-DETR demonstrates a superior balance of accuracy and efficiency, providing a practical solution for deploying high-performance transformer-based detectors in smart agriculture.
No takes yet. Share an insight, caveat, or question.
Xiaorong Zhang (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: