A New Image Reconstruction Algorithm Based On Topic Positive And Negative Perception And Structural Information Mining Residual Network

A super resolution(SR)reconstruction method improves image quality at a lower cost for low resolution signals.It provides more rich details for applications of target detection in aerospace,suburban environment and military missions,urban traffic and security surveilance,biomedical image processing etc.In this dissertation,the structure perception and information mining based SR reconstruction algorithm is studied.And our work is focused on:1.In order to reduce the error between LR and HR matching resulted from the loss of details,a structure-aware SR reconstruction algorithm based on a hidden topic probability model is proposed.Using the statistical results of the natural image set,the topic structures of the LR block are described and reconstructed via a hidden topic probability model.Firstly,a probability model of the hidden topic structure is generated,by assigning the topic prior to the LR image blocks of different content categories.Secondly,the neighborhood context-aware mechanism is proposed to complete the entropy increase process of the LR block information quantity.Then the association information between the neighborhood context and multiple topics are used to adaptively select the LR neighborhood in the manifold structure.Therefore,a differentiated reconstruction for the LR-like signals but with different topics is realized.2.Aiming at the serious loss of structural information after large-scale degradation of HR images,a data driven SR deep reconstruction method is proposed.Varieties of structures are captured by our structurally complementary information mining residual network.Firstly,we propose multi-level residual connections mechanism to combine level-different features globally.It reuses learned shallow features,and forces subsequent feature extraction units to mine lacking information,thereby brining about the complementarity of diverse structure information.Secondly,with the local dense connection strategy,the feature of each residual unit in the residual group is propagated to all subsequent layers locally,achieving continuous propagation and comprehensive utilization of multi-layer features.The constructed dense multi-level residual structure is used as a feature extraction module of the SR depth network,thereby realizing effective recovery of the structural content of the LR signal.