.png){kind=icon}
.png){kind=icon}
Return
Distributed compressed sensing based on local transformer network
.png){kind=icon}
PRE
AI
DOI：10.1016/j.ins.2025.122788.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Distributed Compressed Sensing (DCS) is notable among compressed sensing algorithms for its ability to jointly reconstruct multiple signals. This efficiency makes it a focal point for practical applications. In contrast to conventional approaches that depend on a priori estimates, like joint sparsity of signals, deep neural networks leverage their robust feature extraction and sequence modeling abilities to automatically capture inter-signal dependencies. This confers them with the benefits of faster solving speed and superior reconstruction accuracy. However, most current methods do not fully utilize the global dependencies between signals and often overlook fine-grained intra-signal features. Therefore, this paper proposes a Local Transformer Network for Distributed Compressed Sensing (LTN-DCS). It employs a self-attention mechanism to extract global contextual information from reconstructed signal sequences and addresses local redundancy through partitioning for finer-grained feature capture. This approach generates probabilistic estimates of the non-zero element positions in each signal, which guide the selection of basis atoms from the sampling matrix. Finally, the least-squares method is used to achieve the signal reconstruction. Through extensive experiments on MNIST dataset, brain tumor MRI dataset and natural image dataset, the results show that our proposed method outperforms state-of-the art distributed compressed sensing algorithms, and LTN-DCS has better reconstruction accuracy as well as robustness, and maintains better sparsity. Furthermore, we investigate a sequence modeling approach for training data in distributed compressed sensing. We introduce a Grouping-based LTN-DCS method, GLTN-DCS, to explore the trade-off between training efficiency and reconstruction accuracy. The source code is available at https://github.com/EMRGSZU/papers-code/tree/main/LTN-DCS.
Keywords:
Local transformer
Distributed compressed sensing
Self attention mechanism
Deep learning
Journal
IF:
6.8
Papers:
540
Citations:
6.2W
Organization
View more organizations