Automatic Crop Pest Detection Oriented Multiscale Feature Fusion Approach.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Manual pest monitoring in agriculture is labor-intensive and time-consuming.
• Automated pest detection using deep learning shows promise but faces challenges with small and visually similar pests.
• Existing deep learning models struggle with information loss during training and feature misalignment for small objects.

Purpose of the Study:

• To develop an advanced deep learning network for accurate multi-category pest detection.
• To address the challenges of detecting small pests and distinguishing between visually similar pest categories.
• To improve the efficiency and effectiveness of automated agricultural pest monitoring systems.

Main Methods:

• Proposed the Multi-Category Pest Detection Network (MCPD-net).
• Incorporated a Multiscale Feature Pyramid Network (MFPN) to fuse pest information across different scales.
• Introduced an Adaptive Feature Region Proposal Network (AFRPN) to resolve anchor-feature misalignment, particularly for small pests.

Main Results:

• The MCPD-net achieved a mean average precision (mAP) of 67.3% on the MPD2021 dataset.
• The model demonstrated an average recall (AR) of 89.3%.
• Outperformed existing deep learning-based pest detection models in experimental evaluations.

Conclusions:

• The proposed MCPD-net effectively enhances pest detection accuracy, especially for small and similar-looking pests.
• The MFPN and AFRPN components significantly contribute to improved performance in multi-category pest identification.
• This deep learning approach offers a more efficient and accurate solution for automated agricultural pest monitoring.