Lightweight CNNs for Advanced Bird Species Recognition on the Edge

Adrian Berenguer-Agulló, Javier Rodríguez-Juan, David Ortiz-Perez, José García Rodríguez

Abstract: This study embarked on a comprehensive exploration of deploying lightweight CNN models for the real-time classification of bird species, particularly focusing on their application within edge computing environments. Given the critical importance of rapid species identification in conservation efforts and ecological monitoring, this research aimed to evaluate the balance between model accuracy and computational efficiency. By meticulously selecting metrics such as Raspberry Pi inference time, accuracy on the Birds525 and CUB-200-2011 test sets or model size, we assessed the performance of various pre-trained models. Additionally, we fine-tuned the models to our specific task. Our findings reveal significant insights into the trade-offs between speed and accuracy, highlighting models like EfficientNetV2 (92% accuracy) and MobileNetV3 (0.082 inference time on Raspberry pi) variants as promising candidates for edge-based wildlife monitoring applications. These discoveries underscore the feasibility of employing deep learning techniques in real-time, resource-constrained settings, offering a quality paradigm for ecological research and conservation strategies.