RESE-CNN: Residual Squeeze-and-Excitation Network for High-Contrast Optical Tomography Reconstruction

Abstract

Optical tomography, a critical component of process tomography, is an important tool for determining the absorption coefficient of cross-sectional media, with significant applications in multi-phase flow analysis, chemical processing, and combustion studies. However, the precise reconstruction of these media distributions is a major challenge. In this work, a sophisticated optical tomography (OT) system coupled with an innovative reconstruction algorithm is presented. The system architecture includes 25 light sources and 25 strategically placed fan-beam receivers. We present a convolutional neural network (CNN) with an encode-decode configuration, augmented by residual connections and a squeeze-and-excitation (SE) attention mechanism. Initial evaluations performed using MATLAB simulations showed the algorithm's superior performance compared to existing methods, with notable improvements in relative error (RE) and correlation coefficient metrics. Subsequent practical experiments validated these findings and emphasized the efficiency of the residual and SE components in improving reconstruction accuracy. While this study focuses on high-contrast binary scenarios, the proposed RESE-CNN framework provides a basic architecture for future extensions to weakly absorbing and scattering media where nonlinear reconstruction problems dominate.