PROJECT TITLE :

Learning Converged Propagations With Deep Prior Ensemble for Image Enhancement

ABSTRACT:

Various vision and learning applications rely heavily on improving the visual quality of images. Recently, researchers have looked into both knowledge-driven maximum a posterior (MAP) with prior models and totally data-dependent convolutional neural network (CNN) techniques to tackle specific enhancement challenges. A unified framework, called deep prior ensemble (DPE), is proposed in this research by utilising the advantages of these two types of processes within a complimentary propagation approach. To be more specific, we start with a basic propagation method based on fundamental picture modelling cues and then incorporate residual CNNs to help forecast the propagation direction at each level. With prior projections, we can theoretically verify that even with experience-inspired CNNs, DPE is convergent and its output always meets our fundamental task requirements. Because our descent routes are learned from training data, we have an advantage over standard optimization-based MAP techniques in that they are substantially more resistant to local minimums. DPE, on the other hand, is able to take advantage of the rich task cues examined on the basis of domain knowledge as compared to existing CNN type networks. As a result, the DPE offers a generic ensemble methodology for various image improvement tasks that incorporates both knowledge and data-based signals. Furthermore, our theoretical investigations show that the DPE feed-forward propagations are properly controlled towards our desired solution. On a number of image-enhancing tasks, the suggested DPE beats current methods, both quantitatively and visually, according to the findings of experiments.