Dual-loss path CNNs fusion architecture for image restoration with scattered data

High-resolution scattered data restoration is vital for series analysis of imaging through scattering medium. However, the existence of traditional methods mostly using deterministic transmission matrix, which limits the speckle de-correlations. Thus, learning based networks are applied to do image restoration through different random scattering mediums. In this work, we develop a novel dual-loss path networks fusion architecture for scattered image restoration focusing on improving reconstructed image details and the edge information. Our architecture consists of dual-loss path of convolution neural network (CNN) modules. In particular, we use the perceptual loss to extract the high level feature images, and the objective loss to extract the low level feature images, individually. After the both low level and high level feature manners are reconstructed by dual-path CNNs model, we use our proposed fusion method to combine them in frequency domain. Our network architecture is verified effective on two kinds of public datasets (MINST and Fashion-MINST). The experimental results show that our proposed method achieving better restoration than common CNN methods especially in the edges and the corner information of the predicted images.