An Enhanced Insect Pest Counter Based on Saliency Map and Improved Non-Maximum Suppression

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Accurate insect pest counting from digital images on yellow sticky traps is crucial for effective pest monitoring.
• Existing methods face challenges with small objects and simple backgrounds typical of yellow sticky trap imagery.

Purpose of the Study:

• To develop an improved automated system for counting insect pests from digital images.
• To enhance the accuracy and reliability of insect pest detection and quantification in agricultural settings.

Main Methods:

• A saliency map-based region proposal generator was developed, incorporating saliency map building, activation region formation, and background-foreground classification.
• A convolutional neural network (CNN) model was employed for classifying proposed regions and generating detection bounding boxes.
• An improved non-maximum suppression technique was introduced to refine redundant bounding boxes, followed by counting for pest quantification.
• A dual-path network integrating Faster R-CNN with the proposed counter was constructed to address challenges with closely positioned pests.

Main Results:

• The proposed insect pest counters demonstrated significant improvements in F1 score compared to state-of-the-art object detectors.
• The methods achieved higher accuracy in detecting and counting insect pests, even when individuals were close together.
• The dual-path network effectively handled complex scenarios, leading to more reliable pest counts.

Conclusions:

• The developed saliency map and improved non-maximum suppression approach offers a robust solution for automated insect pest counting.
• The integration with Faster R-CNN further boosts performance, providing a valuable tool for precision agriculture and pest management.
• This work advances automated image analysis techniques for ecological and agricultural monitoring applications.