An EfficientNet-based Ensemble for Bird-Call Recognition with Enhanced Noise Reduction

Abstract Birds are an integral part of the biological ecosystem. However, spotting birds is an arduous task than hearing them chirping. Every bird species has a signature song relying on which ornithologists study birds in an ecosystem. Thus, bird-call recognition plays a pivotal role in identifying birds, understanding biological diversity and conserving them. Previously, a few research works have attempted to devise machine learning and deep learning techniques to identify bird-calls. However, most approaches inadequately focused on pre-processing and eliminating background noise from the audio input. In this paper, a fine-tuned EfficientNet-based ensemble model to classify avian species has been proposed with a novel tri-layered noise reduction approach. It has been augmented with a thresholding-based approach which discards the noisy samples altogether. Furthermore, a tri-point trade-off between the accuracy, model depth and the model parameter has been hypothesized and validated. The proposed methodology surpasses the existing state-of-the-art models yielding an accuracy of 0.97 and an F1-score of 0.95 on the Cornell Birdcall Identification task.