<p>Directly modeling time-domain signals has demonstrated remarkable performance in underwater acoustic target recognition. This approach leverages end-to-end deep learning architectures, eliminating complex manual feature extraction and enabling direct signal-to-category mapping. However, the non-uniform noise distribution and extracted features often lack a clear correlation with acoustic-physical mechanisms, constraining the model’s generalizability and interpretability. To mitigate these issues, this paper proposes a novel Adaptive Spectral Bandpass Multi-scale Network (ASBMNet) that directly learns discriminative target characteristics from raw signals. Specifically, the proposed architecture incorporates a soft frequency masking matrix generated through adaptive thresholding, whose continuity is optimized by smoothing to suppress noise. Furthermore, the reformed Sinc convolution utilizes trainable Sinc functions to learn bandpass filters with transparent frequency responses and traceable learning processes, achieving dynamic adaptation to the input frequency characteristics. To capture multi-scale features, the model employs cascaded convolutional encoders, which enable feature extraction in different spectral ranges by gradually expanding the convolutional kernel scales (short, medium and large). Extensive experiments on two real ocean datasets demonstrate that ASBMNet outperforms current models and exhibits stronger generalization ability under low signal-to-noise ratio scenarios.</p>