国产精品久久久-韩日毛片-丁香六月激情-黄网在线免费观看-毛片网站大全-日韩欧美亚洲国产-巨胸喷奶水wwww贱多-中文字幕av一区二区三区人妻少妇-成人在线中文字幕-囯产精品一品二区三区-一区二区三区四区不卡-午夜涩涩-亚洲你懂得-福利片网址-色呦呦在线免费观看

Underwater acoustic signals from diverse sources exhibit high degrees of overlap across the time, frequency, and spatial domains. Consequently, accurately recovering valuable source components from such complex, aliased mixtures is pivotal for effective underwater acoustic monitoring and analysis. Given the limited resources of underwater sensors, lightweight sound separation models often sacrifice feature depth for efficiency, thereby restricting their capacity to analyze complex acoustic textures. Furthermore, existing sound separation models typically employ traditional static residual connections, which struggle to adapt to the nonstationary nature of the underwater environment. To surmount these limitations, a resource-efficient architecture explicitly tailored for underwater nodes named dynamic gated deep encoder network (DGDE-Net) is proposed. At its core is the weight-sharing deep sub-band encoder (WDSE), which decouples the dependency between parameter count and sub-band resolution, achieving deep nonlinear feature extraction with negligible computational overhead. Additionally, a gated residual connection (GRC) mechanism is introduced to dynamically modulate the weighting between deep feature transformation and raw signal fidelity. Extensive experiments on multiple marine datasets demonstrate that DGDE-Net achieves state-of-the-art (SOTA) performance while maintaining a lightweight footprint of only 1.54 M parameters. Notably, in separating killer whale sounds from passenger ship noise, it attains an SDRi of 11.16 dB and an SISNRi of 11.24 dB. The model exhibits significant robustness in separating complex interspecies interference and biological-vessel noise mixtures, offering a viable solution for high-precision, resource-constrained underwater acoustic sensing.