摘要
提出了一种量化噪声的综合分布模型来估计量化过程带来的误差,并结合深度学习进行超分辨率重建。首先对数据库训练样本进行矫正量化误差的预处理,然后通过训练学习误差矫正后的低分辨率视频帧与高分辨率视频帧之间的特征,获得更加准确的映射关系,减少了量化过程带来的误差和质量损失。实验结果表明,本实验的算法不管在主观体验还是客观指标上都优于以前的算法。
This paper proposes a comprehensive distribution model for quantization noise to estimate the quantitative error and do the super-resolution reconstruction with deep learning.First,we make the preprocessing for all the frames in the training database to correct the quantization error,and then by training features between the low resolution frames and the high resolution frames after error correction,we acquire more accurate mapping relationship,and reduce the error and quality loss brought by quantitative process.Experimental results show that our algorithm is superior to the previous algorithms in both subjective experience and objective index.
作者
王春萌
WANG Chun-meng(Jinling Institute of Technology, Nanjing 211169, China)
出处
《金陵科技学院学报》
2020年第1期10-15,共6页
Journal of Jinling Institute of Technology
基金
金陵科技学院高层次人才科研启动基金(jit-b-201802)
江苏省高等学校自然科学研究面上项目(19KJB520007)。
关键词
超分辨率重建
量化误差
深度学习
卷积神经网络
视频编码
super-resolution reconstruction
quantization error
deep learning
convolutional neural network
video coding
