TD-Det: A Tiny Size Dense Aphid Detection Network under In-Field Environment.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Aphid infestations cause significant crop yield reduction and economic losses.
• Existing aphid detection methods struggle with tiny size, dense distribution, and data quality.
• Real-time detection is essential for effective pest management but often compromised by performance improvements.

Purpose of the Study:

• To develop a robust and efficient aphid detection method addressing limitations of current approaches.
• To improve feature extraction for tiny, clustered pests using customized network designs.
• To achieve real-time detection performance without sacrificing accuracy.

Main Methods:

• Proposed a Transformer feature pyramid network (T-FPN) incorporating feature-wise (FTM) and channel-wise (CFRM) recalibration modules.
• Implemented a multi-resolution training method (MTM) employing a coarse-to-fine training pattern.
• Conducted experiments on a densely clustered tiny pest dataset to validate the TD-Det method.

Main Results:

• The proposed method achieved an average recall of 46.1% and an average precision of 74.2%.
• TD-Det outperformed state-of-the-art methods including ATSS, Cascade R-CNN, FCOS, FoveaBox, and CRA-Net.
• Achieved the fastest training speed and a testing time of 0.045 seconds per image, meeting real-time requirements.

Conclusions:

• The developed TD-Det method accurately and efficiently detects in-field aphids.
• The approach provides a strong foundation for automated aphid detection and ranking systems.
• Customized T-FPN and MTM effectively address challenges in detecting small, densely packed pests.