基于多尺度宽激活残差注意力网络的图像去块效应

为了节约传输带宽和存储资源，成像设备和系统一般对图像和视频进行了有损压缩. 由于分块量化编码，JPEG图像往往存在明显的块效应. 去除图像的块效应不仅能够改善使用者的视觉体验，还有利于其他计算机视觉任务的开展. 为此，本文提出了一种基于多尺度宽激活残差注意力网络（MWRAN）的图像去块效应方法. MWRAN主要由多尺度宽激活残差注意力模块（MWRAB）构建而成. 提出的MWRAB不仅能够激活更多的非线性特征以促进信息在网络中的流动，还能够捕获丰富的图像多尺度特征. 此外，通过提出的轻量的差异感知通道注意力（LCCA），MWRAB能够对学习到的特征进行自适应地调整以关注更重要的信息. 消融实验验证了MWRAB的有效性. 在常用的基准数据集上，MWRAN取得了比几种先进的图像去块效应方法更高的客观评价指标和更接近原图的主观视觉效果.

Image deblocking based on multi-scale wide-activated residual attention network

To save transmission bandwidth and storage resources, imaging devices and systems generally perform lossy compression on images and videos. JPEG images usually suffer from obvious blocking effect due to block quantization coding. Removing the blocking effect of the image can not only improve the visual experience of users, but also facilitate other computer vision tasks. Therefore, an image deblocking method based on multi-scale wide-activated residual attention network (MWRAN) is proposed. The MWRAN is mainly constructed by the multi-scale wide-activated residual attention block (MWRAB). The MWRAB can not only activate more non-linear features to promote the flow of information in the network, but also capture rich image multi-scale features. In addition, the MWRAB can adaptively adjust the learned features to focus on more important information via the proposed lightweight contrast-aware channel attention (LCCA). The ablation experiment is conducted to verify the effectiveness of the proposed MWRAB. The MWRAN achieves higher objective evaluation indices and produces subjective perceptual effects closer to the original image than several state-of-the-art image deblocking methods on common benchmark datasets.