Precise Crop Pest Detection Based on Co-Ordinate-Attention-Based Feature Pyramid Module

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Insect pests significantly impact global crop production and economic value.
• Accurate and rapid pest detection is essential for effective pest management and infestation mitigation.
• Current deep learning methods face challenges in detecting small crop pests due to difficulties in feature extraction and sample selection.

Purpose of the Study:

• To develop an advanced deep learning model for accurate detection and recognition of small-sized crop pests.
• To address the limitations of existing methods in feature extraction and positive/negative sample selection for small pest detection.

Main Methods:

• Designed a co-ordinate-attention-based feature pyramid network (CAFPN) for enhanced salient visual feature extraction.
• Implemented a dynamic sample selection strategy with positive and negative weight functions during network training.
• Evaluated the model on large-scale datasets: AgriPest 21 and IP102.

Main Results:

• The CAFPN model achieved promising detection results on benchmark datasets.
• Achieved mean average precision (mAP) scores of 77.2% on AgriPest 21 and 29.8% on IP102.
• Demonstrated superior performance compared to other existing pest detection models.

Conclusions:

• The proposed CAFPN model effectively overcomes limitations in small pest detection by improving feature extraction and sample selection.
• The dynamic sample selection strategy enhances both classification accuracy and localization precision.
• The results indicate a significant advancement in deep learning-based crop pest detection systems.