Identifying Animals in Camera Trap Images via Neural Architecture Search.

Area of Science:

• Ecological research
• Computer vision
• Edge computing

Background:

• Camera traps generate vast image data crucial for wildlife monitoring and ecological studies.
• Classifying these images traditionally relies on deep convolutional neural networks, which are resource-intensive and difficult to scale.
• Deploying these networks on edge devices for real-time processing presents challenges due to hardware heterogeneity and limited resources.

Purpose of the Study:

• To develop an automated method for designing neural networks suitable for edge devices in camera trap image classification.
• To address the challenges of heterogeneous edge devices and reduce the computational cost of network design.
• To enable scalable and efficient wildlife image analysis in ecological research.

Main Methods:

• Utilized neural architecture search to automate network design for edge devices.
• Employed regression trees to efficiently evaluate candidate networks, reducing search costs.
• Developed a meta-architecture that automatically adjusts to resource constraints of edge devices.

Main Results:

• Successfully identified a suitable network for the Jetson X2 edge device within 6.5 hours.
• The designed network achieved competitive accuracy compared to both automatically and manually designed networks.
• Demonstrated the feasibility of deploying customized networks on heterogeneous edge devices for camera trap image analysis.

Conclusions:

• Automated neural architecture search, combined with regression tree evaluation, provides an efficient solution for designing edge-deployable networks.
• This approach overcomes limitations of traditional deep learning methods for large-scale camera trap data analysis.
• The method enhances the scalability and adaptability of wildlife monitoring systems using edge computing.