基于双域学习的JPEG压缩图像去压缩效应算法

Dual-domain based decompression algorithm for JPEG compressed Images

针对JPEG压缩图像存在的压缩伪影,提出了一种基于双域学习的JPEG压缩图像去压缩效应算法,以使压缩图像达到更好的视觉效果。该算法利用深度卷积神经网络,根据JPEG压缩图像的特点,分别在像素域和DCT变换域对压缩图像进行去噪,最后将双域的学习信息进行有效融合,以达到更好的去块效应效果。所提出的卷积神经网络使用宽激活残差块(Wide-activation Residual Block,WARB)作为结构单元,能在有效提升网络预测性能的同时,不引入更多的网络参数和计算量。实验结果表明,相比于目前先进的去压缩效应算法,所提出的JPEG压缩图像去压缩效应算法能在主客观上均获得更好的性能。

In this paper,a dual-domain based decompression algorithm for JPEG compressed images is proposed.From the essence of distortion caused by JPEG compression,the proposed algorithm uses deep convolutional neural network to remove noise in pixel domain and the DCT domain respectively.Finally,the output information of both domains is fused effectively,so as to realize better deblocking effect.The proposed convolutional neural network uses wide-activation residual block(WARB)as the structural unit,which can effectively improve the prediction performance of the network without introducing more network parameters and calculations.Experimental results show that compared with start-of-the-art decompression algorithms,the proposed decompression algorithm for JPEG compressed images can achieve better performance both subjectively and objectively.