Single Image Super Resolution Reconstruction Algorithm Base On Ateention Nechanism Dense Connection

With the development of artificial intelligence technology,the method of using deep learning method to deal with image restoration problem increasingly become the mainstream Super-resolution reconstruction field as a height direction in the field of image restoration,in recent years,there are many based on the research of deep learning approach;Based on the deep learning method based on a lot while the reconstruction of the excellent results have been achieved,but a lot of method still has its disadvantages.The current mainstream of superresolution algorithm at the time of introducing the attention mechanism,only consider the channel space or attention,also many super-resolution algorithm only within a single space to extract the features,makes features use of inadequate,led to high resolution image reconstruction edge blur,detail is missing,in order to solve the above problem,this master thesis finished two work:1.Under the convolutional neural network framework,based on the idea of dense connectivity,and a blend of spatial attention and channel attention is embedded in the framework of the replacement of attention mechanism,so as to put forward a kind of based on image super resolution reconstruction based on attention mechanism intense connection algorithm,through intensive connection and the features of the model to extract more abundant attention mechanism.2.Based on the above model,the multi-scale thoughts into the network,the first by a number of different scales on the low resolution image feature extraction,then on the basis of the pixel sampling,will be the last and on the characteristics of the deep figure for sampling on the sampling image,to obtain the ultimate reconstruction on the diagram,the improved model can extract the characteristics of multiple scales,make full use of low-level features;After experiments on a wide range of data sets,and compared with the existing super resolution image reconstruction algorithms,the two algorithms proposed in this paper have achieved good reconstruction results.