Detection Method of Citrus Psyllids With Field High-Definition Camera Based on Improved Cascade Region-Based...

Area of Science:

• Agricultural Entomology
• Computer Vision
• Machine Learning

Background:

• Citrus Huanglongbing (HLB) is a devastating citrus disease, primarily transmitted by the citrus psyllid.
• Current manual detection methods for citrus psyllids are labor-intensive and inefficient for orchard management.
• The small size and cryptic coloration of citrus psyllids pose challenges for traditional computer vision algorithms.

Purpose of the Study:

• To develop an accurate and efficient computer vision system for detecting citrus psyllids in natural light conditions.
• To improve the recognition performance of small target pests in citrus orchards.
• To reduce the cost and time associated with manual pest surveillance.

Main Methods:

• Collected a dataset of citrus psyllids and fruit flies using a high-definition camera under natural light.
• Proposed a semantic segmentation-based method to augment small pest data.
• Improved the cascade region-based convolution neural networks (R-CNN) algorithm with multiscale training, CBAM attention, HRNet, sawtooth ASPP, and FPN for enhanced feature extraction and fusion.
• Implemented an online hard sample mining strategy to address difficult detection cases.

Main Results:

• The improved cascade R-CNN model achieved an average recognition accuracy of 88.78% for citrus psyllids.
• The developed algorithm outperformed common networks like VGG16 and ResNet50 in small target recognition.
• The model demonstrated effectiveness in identifying other small targets, such as citrus fruit flies.

Conclusions:

• The enhanced cascade R-CNN model offers a feasible and effective solution for detecting small target pests in citrus orchards using field cameras.
• This computer vision approach has the potential to significantly improve citrus pest management strategies.
• The method's generalization ability under varying outdoor light conditions is a key advantage.