Deep Neural Network Models For Images Super-Resolution

Image s uper-resolution is a process of restoring one or more consecutive low-resolution images of the same scene to high-resolution images by means of hardware or software.High resolution image means that it has higher pixel density,which produces more pixel information and high-frequency details.In practical application,these details often play a key role,and the re is a growing demand in many industries such as medicine,biology,remote sensing,security,film and television.With the development of artificial intelligence in recent years,image super-resolution based on deep learning has broken through the bottleneck of traditional algorithm and achieved better restoration effect.However,the existing models are insufficient to ex-tract the effective features of the image,and they do not make full use of the image context information to reconstruct the texture details,and most of the model parame-ters are large,the calculation is large,and the visual effe ct of the reconstruc ted image is poor.In view of the existing problems and deficiencies in the existing models,this paper proposes two models based on convolutional neural network and generative adversarial network respectively from the perspective of image quality objective metric and visual perception metric.The main contributions are as follows:·In view of the difficulty of deep feature representation,this paper proposes an image super-resolution model based on deep residual dense network and atten-tion mechanism.By removing the BN module in ResNet,global and local residual learning is introduced into the problem of bottom-level visual super-resolution,and dense connection block structure is added to realize the fusion of different levels of features.By using residual learning and dense connection,the depth of network is effectively deepened and the performance of module integration is improved.Furthermore,in order to focus on the effective features in the im-age,according to the mutual dependence between the feature channels,the multi-channel attention mechanism is used,so that the model can fully extract the effec-tive features of the image and improve the overall network performance.Ablation study verified that the proposed residual dense structure and multi-channel atten-tion modules contribute to the improvement of model performance.Comparison tests showed that NSR and SSIM indexes of the model are higher than other mod-els,which proves its validity and accuracy.·Aiming at the problem of overweight model and poor visual perception effect,this paper proposes an image super-resolution model based on generative adversarial network and heterogeneous convolution.In the limited computing and storage resources,it is particularly important to build a lighter basic module.Therefore,using more efficient heterogeneous convolution instead of ordinary convolution can reduce the model parameters without losing the performance of the model,and improve the model performance by building effective residual blacks.Fur-thermore,we use the relativistic generative adversarial network to conduct con-frontation training,and design a visual perception loss function that is more con-sistent with the super-resolution task,so that the task can get better results in visial perception.Ablation study verified that the effective residual block and the heterogeneous convolution have better performanee with less parameters,and the use of relativistic generative adversarial network produces more image details.Compared with other advanced models,the model has better image restoration effect and fewer parameters.