Research On Video And Image Super Resolution Based On Zero-shot Learning

Multimedia content such as images and videos has greatly enriched people's life and entertainment experience;however,the widespread Internet content is limited by some factors such as capture equipment and network bandwidth.Most of them still have low resolution and poor quality,which can hardly meet the current requirements for better image definition.Super-resolution technology can recover the missing detailed information of low-resolution images and videos through algorithms,and generate highresolution images and videos with better visual effects.Therefore,Super-resolution technology has become a research hotspot.Existing deep learning-based super-resolution methods rely on the training process on the external dataset,and most of the low-resolution images used for training are obtained from the ideal degradation model,namely bicubic downsampling,without considering the influence of blur and noise in the real scene.As a result,the performance of these methods degrades dramatically when facing images or videos under non-ideal conditions.The super-resolution method based on zero-shot learning makes full use of the image's internal information and trains only with the low-resolution image to be processed.This kind of method can adapt to super-resolution tasks under multiple degradation models;however,the current method has its limits due to the simple network structure and insufficient prior information,and it is difficult to be directly applied to the field of video super-resolution.Due to these problems,the main research contents of this paper are as follows:1.An improved zero-shot learning image super-resolution method is proposed,combined with the idea of neural network structure designing in the field of superresolution.A versatile convolutional residual block which consists of versatile convolutional layer,deconvolutional layer and residual connection is designed.These modules and jump connections are used to build the overall neural network structure,making the network model more suitable for zero-shot learning super-resolution tasks.Gradient information that conforms to human visual characteristics is introduced,and the gradient feature of low-resolution image is used as additional input to provide more effective prior knowledge for neural network.At the same time,different loss functions are used for different degradation models to train the network in a targeted manner,making the model more applicable.Several comparative experiments are designed to analyze the effectiveness of the network structure,gradient prior,and loss functions in this paper.Experimental results show that the method proposed in this paper performs well in both objective indicators and visual effects.2.Integrate the basis of image super-resolution method,this paper extends the algorithm to the field of video super-resolution.Most video super-resolution methods need to use the previous and following frames' information of the current video frame at the same time during super-resolution,which has some limitations in practical use.This paper defines a video super-resolution model that is more in line with the actual situation,and based on this model,this paper proposes an incremental training video superresolution method which combines gradient features and motion features.This method utilizes multiple features to adaptively determine whether a video frame needs to be added to the training process.Gradient is also used to construct a probability map to determine the key training areas in a video image.The neural network structure is improved for the video super-resolution scene.After the basic network is obtained,the designed feature transformation layer is added to transform the feature map,and the parameters of the feature transformation layer are updated using the fine-tuning training strategy to enhance the performance of the network model.Through experimental analysis,the method proposed in this paper has achieved good performance in video super-resolution.