Deep Learning Based Dual-Modality Fusion for TMR-EMT/ERT

Dual-modality electrical tomography can improve the reconstruction quality of complex media distribution by fusing each single-modality information. The fusion effect of existing dual-modality electrical tomography fusion algorithm based on single-modality reconstructed image is limited, because the inherent nonlinear and ill-posed problems of electrical tomography make the resolution of single-modality images not high along with the loss of information. For the above problem, this paper focuses on directly performing feature processing on measurements of electromagnetic tomography based on tunneling magneto resistance sensor (TMR-EMT) and electrical resistance tomography (ERT), in order to realize the effective fusion of TMR-EMT/ERT dual-modality information. A multi-scale mixed attention network model is proposed to reconstruct the distribution of gas-liquid-solid three phases. The proposed model based on dual-branch structure uses the multi-scale convolution block to directly extract the deep and shallow features of measurement information, and uses the mixed attention block to realize the dual-modality feature fusion in the channel and spatial dimensions simultaneously. Compared with the traditional fusion algorithms, the network model shows better fusion effect and reconstruction quality, and the noise experiment and generalization ability test show that the model has good anti-noise performance and generalization ability.