A hybrid Swin Transformer model for image reconstruction of electrostatic tomography

Abstract

The inverse problem of electrostatic tomography (EST) suffers from limited reconstruction accuracy due to its underdetermined and ill-posed properties. Traditional algorithms and CNN models often exhibit significantly reduced performance in noisy environments. To address this challenge, this study proposes a hybrid model that combines ResNet and Swin Transformer. By integrating the residual connections of ResNet with the multi-scale window attention mechanism of Swin Transformer, the proposed model achieves collaborative optimization for local feature extraction and global dependency modeling. The residual structure of ResNet alleviates the gradient vanishing problem and enhances noise robustness. The window attention mechanism of Swin Transformer enhances the model's ability to capture global features of the charge distribution while reducing computational complexity. By compressing the channel dimension via 1D convolution, the model addresses the feature redundancy problem of Swin Transformer's 2D block partitioning, adapting to the 1D boundary signals of EST. Compared with traditional algorithms, the reconstructed image shows clearer edges and fewer artifacts. The robustness of the model is verified by samples with different signal-to-noise ratios. In addition, random samples are used to verify the generalization ability of the model. The imaging results demonstrate that within the range of 0–50 dB noise, with the increase of noise level, the correlation coefficient decreases and the image error increases. The improved model still maintains higher correlation coefficients and lower image errors under different noise levels, demonstrating good anti-noise performance. Compared with other algorithms, the improved model achieves the highest correlation coefficient (0.9652) and the smallest image error (0.1076), indicating the best imaging performance. The research promotes the practical application of deep learning in solving the EST inverse problem.