基于注意力和多尺寸卷积的超分辨率算法研究

Research on super resolution algorithm based on attention and multi-size convolution

为了解决深层卷积模型的超分辨率技术计算量大、融合的特征不够全面的问题,模型结构不再从深度上进行加深,而是从宽度上进行扩展。对输入的一张特征图进行多尺寸的卷积处理,在结构上融合残差结构、压缩模块和改进的通道注意力模块,融合多尺寸的特征图的同时灵活运用高、低频信息,最终达到提高重建图像质量的效果。实验结果表明:与目前较为流行的超分辨率算法相比,在参数量上有了一定的减少,且在峰值信噪比和结构相似性上有着良好的表现。

In order to solve the problem of large amounts of computation of super-resolution technology of deep layer convolution model, and fusion features are not comprehensive, a model structure is designed no longer deepened in depth, but expanded in width.Multi-size convolution processing is performed on an input feature map, residual structure, compression module and improved channel attention module are fused structurally, and the high-frequency and low-frequency information are flexibly used while fusing the multi-size feature map, and finally achieve the effect of improving the quality of the reconstructed image.The experimental results show that, compared with the currently popular super-resolution algorithm, the amount of parameters are reduced to a certain extent, and it has good performance in peak signal-to-noise ratio and structural similarity.