基于重参数化的超分辨率重建
Super-resolution Reconstruction Based on Re-parameterization
田蕾 1申艺2
作者信息
- 1. 北京跟踪与通信技术研究所 北京 100094
- 2. 南京航空航天大学 南京 210000
- 折叠
摘要
针对现有单图像超分辨率重建(Single Image Super-resolution,SISR)模型存在速度和精度的矛盾,论文给出了一种重参数化(Re-parameterization)的轻量模型用于实现图像重建.该模型训练时通过使用结构较复杂的模型保证精度,推理时通过模型等效变换为简单的卷积以提高速度.同时多监督结构的加入让模型收敛更快且更为灵活.通过峰值信噪比和结构相似度对重建模型的质量和效率进行了评估.验证了所提模型在现有超分辨率重建方法中兼具了轻量和重建质量良好的优点.
Abstract
In view of the contradiction between the speed and accuracy of the existing single image super-resolution(SISR)model,this paper presents a lightweight re-parameterization model for image realization reconstruction.The model is trained to en-sure accuracy by using a model with a more complex structure,and the model is equivalently transformed into a simple convolution to improve the speed during inference.At the same time,the addition of a multi-supervisory structure makes the model converge faster and more flexible.The quality and efficiency of the reconstruction model are evaluated by the peak signal-to-noise ratio and structural similarity.It is verified that the proposed model has the advantages of light weight and good reconstruction quality in the existing super-resolution reconstruction methods.
关键词
单图像超分辨率/卷积神经网络/多监督学习/重参数化Key words
single image super-resolution/convolutional neural network/multi-supervised learning/re-parameterization引用本文复制引用
基金项目
国家自然科学基金项目(61803199)
国家自然科学基金项目(U2033201)
出版年
2024
国家科技期刊平台
NSTL
万方数据