Design And Analysis Of An Exploiting High Level Semantic using Deep Neural Network On Blur Images

Abstract

The human visual system is excellent in detecting local picture fluid, but it is not well known the underlying mechanism. There are fundamental constraints on conventional views of blur, such as energy declines at high frequencies and loss of phase coherence at local features. It is not possible, for example, to separate flat and blurred areas. In this regard, we propose that high-level semantine information should be used to define local blur successfully.We therefore use deep neural networks to research high-level features and suggest the first end-to-end algorithm of local blur mapping based on a totally coherent network. By studying different architectures of different depths and design philosophy, we empirically show that high-level characteristics of deeper layers play a more important role in solving difficult ambiguities in this challenge than features of smaller layers. The proposed system is therefore targeted at Performblur identification, parameter estimation, and deblurring by deep learning in a three- stage system.First, the proposed system uses a supervised method to identify the blur type from a mixed imageinput, i.e. black and white or color image degraded by different blurs with different parametersusing a pre-trained deep neural network (DNN).