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

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Accurate pomegranate maturity detection is vital for optimizing harvest and economic returns.
• Conventional methods struggle with complex agricultural environments, facing limitations in feature representation, attention mechanisms, and multi-scale fusion.

Purpose of the Study:

• To develop an advanced algorithm for precise pomegranate maturity detection.
• To overcome limitations of existing methods in complex agricultural settings.

Main Methods:

• Developed the DFMB-Net backbone for spatial-frequency collaborative processing of pomegranate characteristics.
• Constructed the EAFF module with adaptive sparse attention and multi-scale adapters for robust feature fusion.
• Implemented adaptive interpolation upsampling (AIUP) and multi-branch feature convolution (MFCM) for improved feature alignment and multi-scale representation.

Main Results:

• DFMA-DETR achieved 90.23% mAP@50 and 76.40% mAP@50-95 on the PGSD-5K dataset.
• Demonstrated significant improvements of 3.13% and 3.06% over the baseline RT-DETR model, respectively.
• Showcased superior generalization performance through cross-dataset validation with low model complexity.

Conclusions:

• The DFMA-DETR algorithm offers an effective solution for intelligent pomegranate maturity detection.
• This research advances precision agriculture by enhancing detection technologies.
• The proposed methods improve feature representation, fusion, and multi-scale performance in complex scenarios.