Research And Implementation Of Lightweight Video Super-Resolution Algorithm For End Device

With the increasing improvement of network infrastructure and the widespread popularity of mobile devices in people’s lives,streaming media services and video communication transmission applications have developed rapidly,the demand for using video content on end-side devices is also increasing,and low-resolution video is often difficult to meet user needs,which will seriously affect the dissemination and consumption of video content.Therefore,video super-resolution has become one of the important issues related to end-side devices.In recent years,the video super-resolution algorithm based on deep learning is one of the research hotspots in the field of computer vision,which realizes the mapping from low resolution to high resolution by processing multiple continuous video frame sequences.However,most video super-resolution algorithms tend to build deep neural network models with deep structures and complex connections in order to pursue higher restoration effects.Due to its expensive computational cost and device hardware requirements,the application of video super-resolution technology in daily life is severely limited.In order to complete the deployment of the model on the end-side device and achieve a balance between model performance and effect,this paper proposes a lightweight video super-resolution algorithm for end-side devices.Specifically,this paper proposes a feature extraction module based on the unsharp mask mechanism,which extracts and enhances highfrequency edge features at the feature level,providing more detailed information for subsequent processing.Secondly,this paper proposes a feature fusion module based on dense residual structure to improve the fusion ability of different levels of feature information.In addition,in order to realize the lightweight of the model and reduce the memory consumption and hardware requirements of the model as much as possible,this paper chooses the recurrent neural network structure as the basic architecture of the designed algorithm,and adopts efficient architecture design such as deep separable convolution in the module.In order to realize the end-to-side deployment of the model,this paper adopts methods such as knowledge distillation and model quantization to further reduce parameter memory consumption and ensure a balance between performance and effect.We conducted experiments on public datasets,compared the algorithm proposed in this paper with other methods,and proved that the algorithm in this paper performs well in terms of performance and effect.At the same time,the ablation experiment analysis of each module in the model proves the effectiveness of the proposed model for video super-resolution tasks.Finally,we designed and built a lightweight video super-resolution system deployed on end-side devices,and transplanted the lightweight video super-resolution model to the mobile terminal,the model demonstrates the usability of the system in everyday life by taking a userselected video for super-resolution and returning the result.