In order to solve the problem of the lack of ornamental value and research value of ancient mural paintings due to low resolution and fuzzy texture details,a super resolution(SR)method based on generative adduction ne...In order to solve the problem of the lack of ornamental value and research value of ancient mural paintings due to low resolution and fuzzy texture details,a super resolution(SR)method based on generative adduction network(GAN)was proposed.This method reconstructed the detail texture of mural image better.Firstly,in view of the insufficient utilization of shallow image features,information distillation blocks(IDB)were introduced to extract shallow image features and enhance the output results of the network behind.Secondly,residual dense blocks with residual scaling and feature fusion(RRDB-Fs)were used to extract deep image features,which removed the BN layer in the residual block that affected the quality of image generation,and improved the training speed of the network.Furthermore,local feature fusion and global feature fusion were applied in the generation network,and the features of different levels were merged together adaptively,so that the reconstructed image contained rich details.Finally,in calculating the perceptual loss,the brightness consistency between the reconstructed fresco and the original fresco was enhanced by using the features before activation,while avoiding artificial interference.The experimental results showed that the peak signal-to-noise ratio and structural similarity metrics were improved compared with other algorithms,with an improvement of 0.512 dB-3.016 dB in peak signal-to-noise ratio and 0.009-0.089 in structural similarity,and the proposed method had better visual effects.展开更多
目的在图像超分辨率(super resolution,SR)任务中采用大尺寸的卷积神经网络(convolutional neural network,CNN)可以获得理想的性能,但是会引入大量参数,导致繁重的计算负担,并不适合很多计算资源受限的应用场景。为了解决上述问题,本...目的在图像超分辨率(super resolution,SR)任务中采用大尺寸的卷积神经网络(convolutional neural network,CNN)可以获得理想的性能,但是会引入大量参数,导致繁重的计算负担,并不适合很多计算资源受限的应用场景。为了解决上述问题,本文提出一种基于双阶段信息蒸馏的轻量级网络模型。方法提出一个双阶段带特征补偿的信息蒸馏模块(two-stage feature-compensated information distillation block,TFIDB)。TFIDB采用双阶段、特征补偿的信息蒸馏机制,有选择地提炼关键特征,同时将不同级别的特征进行合并,不仅提高了特征提炼的效率,还能促进网络内信息的流动。同时,TFIDB引入通道关注(channel attention,CA)机制,将经过双阶段信息蒸馏机制提炼的特征进行重要性判别,增强对特征的表达能力。以TFIDB为基础构建模块,提出完整的轻量级网络模型。在提出的网络模型中,设计了信息融合单元(information fusion unit,IFU)。IFU将网络各层级的信息进行有效融合,为最后重建阶段提供准确、丰富的层级信息。结果在5个基准测试集上,在放大倍数为2时,相较于知名的轻量级网络CARN(cascading residual network),本文算法分别获得了0.29 d B、0.08 d B、0.08 d B、0.27 d B和0.42 d B的峰值信噪比(peak singal to noise ratio,PSNR)增益,且模型参数量和乘加运算量明显更少。结论提出的双阶段带补偿的信息蒸馏机制可以有效提升网络模型的效率。将多个TFIDB进行级联,并辅以IFU模块构成的轻量级网络可以在模型尺寸和性能之间达到更好的平衡。展开更多
文摘In order to solve the problem of the lack of ornamental value and research value of ancient mural paintings due to low resolution and fuzzy texture details,a super resolution(SR)method based on generative adduction network(GAN)was proposed.This method reconstructed the detail texture of mural image better.Firstly,in view of the insufficient utilization of shallow image features,information distillation blocks(IDB)were introduced to extract shallow image features and enhance the output results of the network behind.Secondly,residual dense blocks with residual scaling and feature fusion(RRDB-Fs)were used to extract deep image features,which removed the BN layer in the residual block that affected the quality of image generation,and improved the training speed of the network.Furthermore,local feature fusion and global feature fusion were applied in the generation network,and the features of different levels were merged together adaptively,so that the reconstructed image contained rich details.Finally,in calculating the perceptual loss,the brightness consistency between the reconstructed fresco and the original fresco was enhanced by using the features before activation,while avoiding artificial interference.The experimental results showed that the peak signal-to-noise ratio and structural similarity metrics were improved compared with other algorithms,with an improvement of 0.512 dB-3.016 dB in peak signal-to-noise ratio and 0.009-0.089 in structural similarity,and the proposed method had better visual effects.
文摘目的在图像超分辨率(super resolution,SR)任务中采用大尺寸的卷积神经网络(convolutional neural network,CNN)可以获得理想的性能,但是会引入大量参数,导致繁重的计算负担,并不适合很多计算资源受限的应用场景。为了解决上述问题,本文提出一种基于双阶段信息蒸馏的轻量级网络模型。方法提出一个双阶段带特征补偿的信息蒸馏模块(two-stage feature-compensated information distillation block,TFIDB)。TFIDB采用双阶段、特征补偿的信息蒸馏机制,有选择地提炼关键特征,同时将不同级别的特征进行合并,不仅提高了特征提炼的效率,还能促进网络内信息的流动。同时,TFIDB引入通道关注(channel attention,CA)机制,将经过双阶段信息蒸馏机制提炼的特征进行重要性判别,增强对特征的表达能力。以TFIDB为基础构建模块,提出完整的轻量级网络模型。在提出的网络模型中,设计了信息融合单元(information fusion unit,IFU)。IFU将网络各层级的信息进行有效融合,为最后重建阶段提供准确、丰富的层级信息。结果在5个基准测试集上,在放大倍数为2时,相较于知名的轻量级网络CARN(cascading residual network),本文算法分别获得了0.29 d B、0.08 d B、0.08 d B、0.27 d B和0.42 d B的峰值信噪比(peak singal to noise ratio,PSNR)增益,且模型参数量和乘加运算量明显更少。结论提出的双阶段带补偿的信息蒸馏机制可以有效提升网络模型的效率。将多个TFIDB进行级联,并辅以IFU模块构成的轻量级网络可以在模型尺寸和性能之间达到更好的平衡。
