HybridDenseU-Net: learning a multi-scale convolution and dense connectivity CNN for inverse imaging problems

Abstract

Inverse imaging problems (IIPs) is a cutting-edge technology which is part of the nonlinear inverse problem, the solution approaches to which have placedattention on deep learning recently. This paper proposes a unique learning-based framework for IIPs, referred to as HybridDenseU-Net, which takes U-Net as the backbone and optimizes the encoder as a two-branch feature extraction module. Compared to the direct skip-connection in conventional U-Net, dense connections are introduced to merge features between feature maps with the same dimension and construct multi-scale content in the decoder. The validation of HybridDenseU-Net is carried out by a case study of electrical impedance tomography, which is of typical nonlinear IIP. The results illustrate that HybridDenseU-Net has root mean square error of 3.0867 and structural similarity of 0.9846, which are significantly better than some state-of-the-art deep learning-based frameworks. It has been proven that this work could provide a promising idea for future research on learning-based image reconstruction methods.