摘要
目前卷积神经网络已经成为视频超分辨率重构的主流方法。针对目前方法存在着细节信息恢复效果不理想、感知质量差的缺点,提出了一种结合注意力机制的残差卷积网络。通过一维通道注意力与二维空间注意力增强目标物体特征映射,完善重构图像的细节信息,并为单一像素计算一维分离卷积核,显著减少网络计算负担。采用特征损失与像素损失组合函数来训练神经网络以产生高质量的视频帧。实验结果表明,通过PSNR,SSIM和感知距离将所提模型与当前最先进的模型进行比较,该模型能够获得最优的感知距离,重构图像拥有较高的感知质量。
At present,convolutional neural networks have become the mainstream method for video super-resolution reconstruction.Aiming at the shortcomings of the current method,the recovery effect of detailed information is not ideal;and the perceived quality is poor.In this paper,a residual convolution network that combines the attention mechanism is proposed.The one-dimensional channel attention degree and two-dimensional space attention degree are used to enhance the target object feature map;and the detailed information of the reconstructed image is improved;and the one-dimensional separated convolution kernel is calculated for a single pixel,which significantly reduces the network computing burden.A neural network is trained by using a combination of feature loss and pixel loss to produce high quality video frames.The experimental results show that the model is compared with the current state-of-the-art model by PSNR,SSIM and perceptual distance.The model can obtain the optimal perceptual distance and the reconstructed image has higher perceptual quality.
作者
董猛
吴戈
曹洪玉
景文博
于洪洋
DONG Meng;WU Ge;CAO Hong-yu;JING Wen-bo;YU Hong-yang(School of Electronic Information and Engineering,Changchun University of Science and Technology,Changchun 130022;School of Optoelectronic Engineering,Changchun University of Science and Technology,Changchun 130022)
出处
《长春理工大学学报（自然科学版）》
2020年第1期82-88,共7页
Journal of Changchun University of Science and Technology(Natural Science Edition)
基金
吉林省科技厅项目(20160204009GX,20170623004TC)
国家科技部项目(2018YFB1107600)。
关键词
注意力机制
超分辨率
分离卷积
特征损失
attention mechanism
super resolution
separable convolution
feature loss
