Hardware acceleration of electromagnetic tomography image reconstruction based on FPGA

Electromagnetic tomography faces significant challenges due to its ill-posed and nonlinear inverse problem, which impairs image reconstruction quality and increases computational cost. This paper proposes an efficient deep learning-based image reconstruction method, accelerated by a customized convolutional neural network implemented on FPGA, where traditional fully connected layers are replaced with convolutional layers. A high-quality dataset was generated through joint simulation with COMSOL and MATLAB to train the model. Convolution and pooling operations were implemented as hardware IP cores via high-level synthesis, ensuring efficient execution on FPGA’s programmable logic. The design was implemented and validated on Xilinx Zynq-7000 system-on-chip. Experimental results show a 30.5% reduction in execution time compared to an ARM-based implementation, while achieving high reconstruction accuracy with an average relative error of 0.4503 and a correlation coefficient of 0.8632. These results highlight the potential of the proposed method for enabling real-time and online imaging in practical electromagnetic tomography applications.