Prior knowledge auxiliary for few-shot pest detection in the wild

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Deep learning excels at pest detection but struggles with rare species and limited datasets.
• Current methods are restricted to common pests in controlled settings, hindering broader application.
• Large labeled datasets are a bottleneck for training and fine-tuning deep convolutional neural networks for pest detection.

Purpose of the Study:

• To develop a few-shot pest detection network for identifying rare pest species in natural environments.
• To address the limitations of existing deep learning models in pest detection.
• To improve smart plant protection strategies for diverse agricultural and forestry contexts.

Main Methods:

• Introduction of a prior-knowledge auxiliary architecture for few-shot pest detection in the wild.
• Creation of a hierarchical few-shot pest detection dataset collected in natural environments in China.
• Proposal of a pest ontology relation module integrating insect taxonomy and inter-image similarity.

Main Results:

• The proposed few-shot pest detection network demonstrated comparable performance to existing algorithms.
• The model achieved competitive results in terms of mean average precision (mAP) and mean average recall (mAR).
• Experimental validation confirmed the effectiveness of the novel few-shot detection architecture.

Conclusions:

• The developed few-shot pest detection network shows significant promise for identifying rare pest species.
• The approach effectively overcomes the data limitations of traditional deep learning methods.
• This advancement contributes to more robust and adaptable smart plant protection systems.