A study on edge devices for image classification of the Tasmanian devil (Sarcophilus harrisii) for vaccine delivery

A target-specific bait dispenser is required for oral bait vaccination of the endangered Tasmanian devil ( Sarcophilus harrisii ) against the devil facial tumour disease. Development of a dispenser that can be extended to other species would be a beneficial conservation tool. A camera-based edge device with an onboard deep learning model can be used to identify the target species using image classification. 2. It is important to choose a suitable edge device for the dispenser's effective application in power constrained environments. In this study, we evaluated four edge devices for the smart bait dispenser— ArduinoPro Nicla Vision, ArduinoPro Portenta H7 with Vision Shield (LoRa), Raspberry Pi 3 Model B+ and Raspberry Pi Zero 2 W. Two simple convolutional neural networks (CNN1 and 2), and four fine-tuned pretrained models (MobileNetV2, MobileNetV3Small, ResNet50V2, and ResNet152V2) were trained on trail camera images of devil and non-devil species. These models were evaluated across four metrics and post-training quantised for deployment. The edge devices were assessed on inference latency of each model in seconds (s) and power consumption in watt (W). 3. We found that a simple CNN based image classification model yielded the best overall result of ≈96% across all metrics. Raspberry Pi 3 Model B+ and Raspberry Pi Zero 2 W could run all the six models whereas Portenta failed to run the two ResNet models and Nicla Vision failed to run all four fine-tuned pretrained models. All edge devices were found to be quick enough for dispenser application (average inference latency of 0.718 s). Portenta consumed the lowest power during inference (0.19 W), idle (0.185 W), and light sleep (0.045 W) states whereas Nicla Vision consumed the lowest power during deep sleep state (0.002 W). 4. Overall, we conclude that a simple CNN image classifier, CNN1, deployed on Portenta is the most suitable edgeAI application for a smart bait dispenser. We suggest integration of this system with a sleep-mode efficient microcontroller unit like ATtiny85 for multiday field application. This method can be readily adapted for other species, such as eastern quoll and racoons, for projects like supplemental feeding and vaccination.