Automated detection of 20 Hz fin whale calls using computer vision techniques

Efficient processing of long-term acoustic recordings is essential to study marine species, as manually analysing hundreds of hours of data is time-consuming and prone to fatigue. In this work, we present a simple but effective method for detecting fin whale (Balaenoptera physalus) 20 Hz calls based on classical computer vision techniques applied to spectrograms and compare it with Deep Neural Networks (DNNs) and a time-domain envelope detector. Acoustic signals are converted into spectrograms via the Fourier transform, enhanced with image filtering, and analysed using a circular Hough transform to detect calls. In a subset of hand-labelled recordings, the method achieved 87.3% accuracy and 80% precision, outperforming an approach based on Deep Neural Networks. When applied to three months of Mediterranean recordings, our detections showed a high correlation (0.95) with expert annotations, successfully capturing periods of peak whale presence. The method was further tested on a publicly available Antarctic dataset, yielding qualitatively comparable results to previously published time-domain analyses. Despite its simplicity, the proposed method is computationally efficient, does not require training data, and can be easily tuned, making it suitable for real-time monitoring and accessible to non-specialists. These results demonstrate that classical image-based techniques can provide robust detection of 20 Hz fin whale calls and support ecological acoustic monitoring effectively. • We present a simple and effective automated method for detecting 20 Hz fin whale calls using classical image processing and pattern recognition applied to spectrograms. • The specialized, morphology-driven approach achieves high correlation with manual ground truth (Mediterranean dataset) and outperforms generic methods, including Deep Neural Networks, in peak detection and computational efficiency. • Our method is computationally lightweight, suitable for real-time monitoring on resource-limited platforms, and does not require training data. • The method generalises reasonably well to Antarctic datasets, providing a visual, morphology-driven framework for passive acoustic monitoring. • All data and code used in the study are publicly available, ensuring full reproducibility of results.