Improvement Of Deep Learning Method For Single Image Super-resolution

Single image super-resolution aims to generate a visually pleasing high-resolution image from a low-resolution image.Compared to conventional methods,convolutional neural network based methods have achieved great success.However,previous convolutional neural network structures will encounter model degradation problem when construct a deeper model,which leads to decl:ine of super-resolution performance.In addition,mean square error suffering from some limitations when regarded as a measure of image signal quality was used as loss function in the most convolutional neural network based image super-resolution methods.In respect of the issues above,we propose two improvement methods in this work respectively.1.Densely residual network and local feature fusion based single image super-resolution method.Conventional single path convolutional neural network structure will encounter model degradation problem when construct a deeper model,as well as the network structure with single global residual.Firstly,the proposed densely residual network can optimum model effectively,because it make use of the advantage of global residual and extend the residual prediction to every convolutional layer.Secondly,we bring the local feature fusion into the densely residual network.The modified network can take full advantage of the features learned in the shallow layers and improve the expression ability of local features.The experimental results demonstrate that the densely residual network has a better super-resolution effect on the standard dataset.2.l₁ and structural similarity error loss function based convolutional neural network for single image super-resolution method.The conventional convolutional neural network based image super-resolution mostly uses the mean square error as loss function.However,in image field the mean square error does not satisfy the human perception of image quality,that is,it cannot capture the internal characteristics of the human visual system.Structural similarity index is designed based on the sensitivity of the human visual system to the local structure and is a differentiable function.Moreover,combining l1 error can effectively overcome the shortcoming that structural similarity index is not sensitive to uniform biases.The experimental results demonstrate that the mix loss function has a better super-resolution effect on the standard dataset compared to mean square error,l1 error and structural similarity error loss function.