基于变换域注意力机制的压缩伪影去除的方法

Compression artifacts removal method based on transform domain attention mechanism

图像有损压缩过程往往会导致图像质量退化,使图像出现压缩伪影。针对现有基于深度学习的方法缺乏对联合图像专家组(Joint Photographic Experts Group,JPEG)压缩算法先验信息的利用,提出一种基于变换域注意力机制的去伪影方法。该方法利用卷积神经网络在像素域和离散余弦变换(discrete cosine transform,DCT)域分别提取特征,再将双域学习的特征信息进行融合。利用量化表设计了DCT注意力机制,该模块根据DCT系数的损失程度给予各频率系数不同的权值,使网络自适应补偿量化引起的误差。于此基础上,在像素域引入通道注意力机制,从而更好地利用量化表的先验信息。在主要数据集上,提出的去伪影方法以固定的模型参数对多种质量因子的压缩图像进行伪影去除实验。实验结果表明,所提出的方法在各评价指标和主观视觉上取得较好的效果。

The process of lossy image compression often leads to image quality degradation and compression artifacts.Aiming at the lack of using the prior information of the joint photographic experts group(JPEG)compression algorithm in the existing methods based on depth learning,this paper proposes a compression artifacts removal method based on transform domain attention mechanism.This method uses convolution neural network(CNN)to extract feature information in pixel domain and discrete cosine transformation(DCT)domain.Then,we combine the feature information of dual-domain network.This paper designs a DCT attention mechanism and the module can give different weights to each frequency coefficient according to the loss degree of DCT coefficients enabling the network to adaptively compensate for errors caused by quantization.Furthermore,we use channel attention mechanism in the pixel domain network to make better use of the prior information.The experimental results on LIVE1 and BSDS500 show that,the proposed method has achieved good results in various evaluation indicators and subjective vision.