MSR-RCNN: A Multi-Class Crop Pest Detection Network Based on a Multi-Scale Super-Resolution Feature Enhancement...

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Pest disasters significantly impact crop yields, and accurate pest recognition remains a critical challenge.
• Existing pest detection methods often fail to account for pest distribution and positional requirements, limiting their effectiveness.
• Developing advanced automated systems is crucial for mitigating crop losses due to pests.

Purpose of the Study:

• To propose a robust pest detection network addressing limitations in current methods.
• To enhance the detection of small-size, multi-scale, and high-similarity pests.
• To introduce a large-scale annotated dataset for agricultural pest detection research.

Main Methods:

• A novel pest detection network incorporating a multi-scale super-resolution (MSR) feature enhancement module.
• Integration of a Soft-IoU (SI) mechanism to refine position-based detection accuracy.
• Development and utilization of the large-scale light-trap pest dataset (LLPD-26) with 26 pest classes.

Main Results:

• The proposed method achieved a mean Average Precision (mAP) of 67.4% on the LLPD-26 dataset.
• Demonstrated significant performance gains of 15.0% and 2.7% over state-of-the-art methods AF-RCNN and HGLA, respectively.
• Ablation studies confirmed the effectiveness of the MSR and SI components.

Conclusions:

• The developed pest detection network offers a significant advancement in identifying agricultural pests.
• The MSR and SI mechanisms effectively improve detection accuracy, particularly for challenging pest characteristics.
• The LLPD-26 dataset provides a valuable resource for future research in automated agricultural pest detection.