Image reconstruction of electrostatic tomography based on the improved residual network

Area of Science:

• Electrical Engineering
• Computer Vision
• Applied Physics

Background:

• Electrostatic tomography (EST) relies on solving inverse problems, which are challenging due to limited measurement data compared to the required reconstruction detail.
• Traditional EST methods struggle with the ill-posed nature of the inverse problem, leading to significant reconstruction difficulties.

Purpose of the Study:

• To develop an advanced deep learning model for accurate and robust image reconstruction in electrostatic tomography.
• To enhance the nonlinear expression and generalization capabilities of EST reconstruction models.

Main Methods:

• An improved ResNet-34 network (P-ResNet) was designed, featuring specific residual block configurations (3, 4, 4, 3) and incorporating ReLU activation after the second convolution in each residual block.
• L2 regularization loss function was introduced to improve model generalization.
• Extensive simulations (15,930 samples) and experiments using a novel small sensor for induced charge measurement were conducted.

Main Results:

• The P-ResNet model achieved high accuracy in EST image reconstruction tasks after 200 iterations.
• The model demonstrated significant anti-noise ability when tested with varying levels of Gaussian white noise.
• Image correlation coefficients were higher compared to traditional algorithms, and experimental results validated simulation findings.

Conclusions:

• The proposed P-ResNet model offers a highly effective and generalizable solution for electrostatic tomography image reconstruction.
• The integration of enhanced network architecture and regularization techniques significantly improves reconstruction performance and robustness against noise.