基于增强隐式神经表示的图像超分辨算法研究

Image super-resolution based on enhanced implicit neural representation

隐式神经表示为数字图像的连续表示提供了一种方法,该方法已成功应用于图像超分辨任务中,并能够取得良好的性能。但是,由于其像素级采样的插值策略,导致权重分配失衡,使得恢复出的高分辨率图像边缘、纹理过平滑;同时由深度网络提取的特征图存在底层特征失真的问题。针对上述问题,本文提出一种基于增强隐式神经表示的图像超分辨重构算法(WCESR)。方法中引入权重修正模块,学习局部面积权重与全局结构权重的关系,缓解权重分配失衡现象;同时引入低分辨图像的边缘特征,扩展由深度神经网络得到的深层图像特征,产生锐利的边缘。通过大量对比实验和消融实验证明:本方法可以得到与现有算法相当甚至更好的效果。

Implicit neural representation provides a method for continuous representation of digital images.This method has been successfully applied to image super-resolution tasks and can achieve good performance.However,due to the imbalance of weight distribution caused by the interpolation strategy of pixel-level sampling,the edge and texture of the restored high-resolution image are too smooth;At the same time,the feature map extracted from the depth network has the problem of distortion of the underlying features.To solve these problems,this paper proposes an image super-resolution reconstruction algorithm based on enhanced implicit neural representation,called WCESR.This method introduces the weight correction module to learn the relationship between the local area weight and the global structure weight,and alleviate the imbalance of weight distribution.At the same time,the edge feature of low resolution image is introduced to extend the deep image feature obtained by depth neural network to produce sharp edges.Through a large number of comparative experiments and ablation experiments,it is proved that this method can get equivalent or even better results than the existing algorithms.