Towards accurate bird sound recognition through multi-scale texture-aware modeling

Bird sound recognition poses challenges due to complex, overlapping spectral patterns. We propose a novel framework that combines multi-scale texture-aware modeling with interpretable deep learning. Central to our method is the Directional Laplacian of Gaussian Network (DLoGNet), a convolutional architecture with learnable orientation and scale parameters to capture directional acoustic textures. Additionally, we design the Frequency Band Recalibrated Spectrogram (FBRS), which adaptively selects energy-dense sub-bands via wavelet packet decomposition. Experiments on real-world datasets show that our method outperforms conventional CNNs, RNNs, and attention-based models in both accuracy and class separability. Visualizations of learned filters and t-SNE embeddings support its interpretability and effectiveness. This study highlights the importance of directional and multi-scale features in acoustic signal understanding and offers a robust solution grounded in the principles of explainable artificial intelligence (XAI), providing interpretable directional features and visual insights into model decisions for bird species identification.