基于双注意力机制的轻量级图像超分辨率重建

LIGHTWEIGHT IMAGE SUPER-RESOLUTION RECONSTRUCTION BASED ON DUAL ATTENTION MECHANISM

目前许多超分辨率重建算法不断深化和拓宽网络来提高性能,导致实际应用中计算复杂度和内存消耗增加;同时在深度网络中对特征不加区分地向后传递,导致无法很好地重建高频信息。为此提出一种基于双注意力机制的轻量级图像超分辨率重建算法。将原始低分辨率图像为输入;在网络部分引入双注意力机制增强模块放在模型首尾,分别对浅层和深层特征进行提取和动态校准;模型中部叠加轻量化特征提取模块逐步细化浅层特征;把重建模块上采样到目标尺寸后的图像与插值到目标尺寸的输入相加输出结果。通过在基准数据集上进行实验表明,在仅采用基础模型(VDSR)2/3参数量的情况下平均提高了0.26 dB峰值信噪比,且视觉效果最好。

At present,many super-resolution reconstruction algorithms continue to deepen and broaden the network to improve performance,resulting in increased computational complexity and memory consumption in practical applications.At the same time,the indiscriminate backward transfer of features in the deep network results in the inability to achieve good reconstruction.For this reason,the paper proposes a lightweight image super-resolution reconstruction algorithm based on dual attention mechanism.The original low-resolution image was taken as input.The dual attention mechanism enhancement module was introduced at the beginning and the end of the model to extract and dynamically calibrate the shallow and deep features.The lightweight feature extraction module was gradually superimposed in the middle of the model to refine the shallow features.The image after sampling up to the target size by the reconstruction module was added to the input interpolated to the target size to output the result.Experiments on the benchmark data set show that the peak signal-to-noise ratio is improved by 0.26 dB on average when only 2/3 of the basic model(VDSR)is used,and the visual effect is the best.