Research On Super-Resolution Reconstruction Algorithm Based On Convolutional Neural Network

As the main medium of information,images play a vital role in many fields,and people's demand for high-quality images is becoming more and more urgent.Super-resolution reconstruction technology can break through the limitations of hardware devices,and use low-resolution images to reconstruct high-quality images with higher resolution,richer texture details and clearer images through a series of algorithms.With the rapid development of deep learning,convolutional neural networks have been widely used in super-resolution reconstruction algorithms and have achieved good results.Therefore,this paper will focus on studying and improving image super-resolution reconstruction algorithms based on convolutional neural networks.On the basis of introducing the classical super-resolution reconstruction algorithm based on deep recursive convolution network,this paper improved a super-resolution reconstruction algorithm based on recursive residual network to overcome the difficulty of training deep network model.The specific methods of the algorithm used both global residual learning and local residual learning strategies in the recursive network model.Global residual learning estimated the residual function between the input and output images of the whole network,while local residual learning estimated the residual function between the input and output images of the intermediate convolution layer,which retaind a large number of image details and help to restore high-frequency texture information.The recursive block structure with residual units was used to increase the depth of the network without adding the number of parameters,which made the training difficulty of the network model significantly reduced.At the same time,the reconstruction speed of the network model was accelerated,and the subjective visual effect and objective evaluation index of the reconstructed image were both effectively improved.This paper studied the model structure and the perceptual loss function of thegenerative adversarial network in detail.It also analyzed and summarized the deficiencies in the network,and the corresponding improvements were made in the structure of residual blocks in the generator network,discriminant method of the discriminator network and the location of the perceptual loss function.Then the large dataset ImageNet was used to pre-train the existing excellent networks,and transfer learning was used to transfer the optimal parameters to the improved generative adversarial network.Experiments showed that the algorithm effectively removed artifacts and obtained a clearer and more realistic reconstructed image in subjective vision.