文献类型：期刊文献

英文题名：Improved Wave-U-Net Blind Source Separation Using Content-Aware Filtering and Dynamic Downsampling

作者：Tang, Sainan Long, Fei Hu, Shengbo Liu, Quan

第一作者：Tang, Sainan

通信作者：Long, F[1]|[144401c07006f9bfbbbd7]龙飞;[14440e98202afd796d385]龙飞;

机构：[1]Guizhou Normal Univ, Sch Big Data & Comp Sci, Guiyang 550025, Peoples R China;[2]Guizhou Inst Technol, Coll Artificial Intelligence & Elect Engn, Guiyang 550003, Peoples R China

第一机构：Guizhou Normal Univ, Sch Big Data & Comp Sci, Guiyang 550025, Peoples R China

通信机构：corresponding author), Guizhou Inst Technol, Coll Artificial Intelligence & Elect Engn, Guiyang 550003, Peoples R China.|贵州理工学院;

年份：2025

卷号：33

期号：10

起止页码：3974-3984

外文期刊名：ENGINEERING LETTERS

收录：EI(收录号：20254519455581);Scopus(收录号：2-s2.0-105020793029);WOS:【ESCI(收录号:WOS:001591149200008)】；

语种：英文

外文关键词：Blind Source Separation; decimation; content-aware filter; dynamic downsampling

摘要：This paper proposes Wave-U-Net+CAFDDS, an enhanced variant of Wave-U-Net that improves the performance of Blind Source Separation. To mitigate aliasing and feature loss in Wave-U-Net's decimation layers, we introduce two key enhancements. First, we replace all decimation layers with dynamic downsampling (DDS) layers. DDS adaptively selects sampling positions based on input features, thereby enhancing the model's ability to retain important features. Second, we insert a content-aware filter (CAF) before each downsampling stage. The CAF dynamically modulates its parameters according to feature context, reducing aliasing artifacts and boosting separation quality. We assess our model on the MUSDB18HQ dataset. Experimental results demonstrate that Wave-U-Net+CAFDDS, integrating both CAF and DDS, significantly outperforms the original Wave-U-Net.