| As a challenging ill-posed problem,image super-resolution has always been a research hotspot in the field of computer vision,and has great application prospects in many fields of social production and life.Benefiting from the substantial improvement in computer hardware computing power and the advancement of algorithms,deep learning algorithms have achieved optimal performance in the field of image super-resolution,but deep learning algorithms still face many challenges in real scenarios.Therefore,this thesis focuses on image super-resolution algorithms in real scenes.By designing more complex neural network architectures and selecting more effective optimization objectives,the performance of deep learning-based image superresolution algorithms has been continuously improved.Due to the poor generalization performance of the current algorithms,the performance will drop sharply when dealing with complex and unknown degradation problems in real scenarios.Aiming at the above problem,this thesis proposes a spatially variant degradation image super-resolution network.The spatially variant blur kernel estimation module provides a pixel-wise spatially variant blur kernel estimation for the super-resolution network,so that the network can not only deal with the problem of different degradations between different images,but also solve the problem of inconsistent degradations in different regions within the same image.In addition,in order to make full use of the degradation information provided by the blur kernel estimation module,an affine transformation layer is added to each residual block of the network to adjust the feature parameters in the network,thereby obtaining a higher quality reconstructed image.Experimental results show that the proposed algorithm outperforms the classical SR algorithms when dealing with complex and unknown degradation problems in real scenarios.The pixel-wise spatially variant blur kernel estimation of the image endows the network with a strong ability to deal with different degradations,effectively solving the complex and unknown degradation problems in real scenes.However,explicit pixel-wise blur kernel estimation increases the complexity of the network to a certain extent.In response to the above problem,this thesis employs unsupervised contrastive learning to train an encoder to learn abstract degradation representations of images,and takes the abstract features as the basis for the network to identify different degradation situations.In order to make full use of the degradation representation information,an adaptive feature modification layer is adopted in each residual block of the network to adjust the feature parameters in the network.The experimental results show that the quantitative PSNR has a performance improvement of up to 9% compared with the traditional algorithms on different test sets and the subjective visual performance is also better than the traditional SR algorithms while reducing the network complexity. |
