DFMA-DETR: a pomegranate maturity detection algorithm based on dual-domain feature modulation and enhanced attention

TL;DR

This paper introduces DFMA-DETR, a new algorithm for detecting pomegranate maturity using advanced image processing techniques to improve accuracy in agriculture.

Contribution

The novel DFMA-DETR algorithm introduces a backbone network, enhanced attention fusion, and modules for better feature representation and upsampling.

Findings

DFMA-DETR achieves 90.23% mAP@50 and 76.40% mAP@50-95 on the PGSD-5K dataset.

It outperforms the baseline RT-DETR model by 3.13% and 3.06% while maintaining low complexity.

The algorithm shows strong generalization performance through cross-dataset validation.

Abstract

Accurate detection of pomegranate maturity plays a crucial role in optimizing harvesting decisions and enhancing economic benefits. Conventional approaches encounter significant challenges in complex agricultural scenarios, including limited feature representation capabilities, singular attention mechanisms, and insufficient multi-scale information fusion. This study presents the DFMA-DETR algorithm, which establishes an end-to-end detection framework through dual-domain feature modulation and enhanced attention mechanisms. The core contributions include: (1) Development of the DFMB-Net backbone network that employs spatial-frequency collaborative processing to model pomegranate surface textures, color variations, and morphological characteristics. (2) Construction of the EAFF enhanced attention feature fusion module that integrates adaptive sparse attention mechanisms with multi-scale feature adapters, effectively addressing feature representation challenges under complex background interference; (3) Introduction of the AIUP adaptive interpolation upsampling processor and MFCM multi-branch feature convolution module, substantially improving feature alignment accuracy and multi-scale representation performance. Experimental validation on the constructed PGSD-5K dataset demonstrates that DFMA-DETR achieves detection accuracies of 90.23% mAP@50 and 76.40% mAP@50-95, representing improvements of 3.13% and 3.06% respectively over the baseline RT-DETR model, while maintaining relatively low model complexity. Cross-dataset validation further confirms the superior generalization performance of the proposed approach. This research provides an effective solution for advancing intelligent detection technologies in precision agriculture.

Figures