A Novel Network Learning for Image Compressive Sensing

Abstract

AbstractMost traditional algorithms suffer from intensive computation for compressive sensing image reconstruction. In recent times, algorithms of deep learning-based reconstruction have been proposed to minimize the complexities of time rather than algorithms of iterative reconstruction. However, the deep learning networks’ training phase has to reduce the complexity of the time. An innovative selective deep residual reconstruction network (SDRRN) learning algorithm is proposed to recreate the image from its measurement of compressively sensed (CS). The SDRRN contains two components: linear mapping of the network and residual networking. Simple fully interconnected layers are incorporated in the linear mapping network. A novel learning strategy is proposed which only propagates the selected training components in the residual learning process. Extensive experimentation shows that the SDRRN outperforms conventional iterative methods and also low-time complexity when compared with other deep learning-based methods.KeywordsCompressive sensingDeep learningReconstructionConvolution neural networkResidual learningNetwork mapping