多尺度信息蒸馏的轻量级图像超分辨率算法

Super-resolution algorithm for lightweight image based on multi-scale information distillation

针对现有的图像超分辨率算法存在细节信息恢复能力较弱、特征复用不合理的问题,提出一种结合信息蒸馏及双链路上采样的超分辨率重建算法.首先,通过多尺度信息蒸馏模块对特征进行多维度提取,使获取的特征信息更全面,增强网络的表征能力;其次,蒸馏机制将多尺度特征进行选择性提炼,并将蒸馏出的部分特征利用层次注意力机制进行全局复用,不仅降低了网络参数,还能获取更丰富的上下文信息;最后,对不同路径下获取的特征分别上采样,将局部和全局特征相结合,提高对细节信息恢复的能力.实验结果表明,所提算法重建出的图像质量更佳,在4倍放大系数下的平均峰值信噪比值比特征蒸馏交互加权网络(FDIWN)提升了0.35 dB,模型参数量相对于级联残差网络(CARN)降低了55%,其性能超过了当前主流轻量级算法.

Aiming at problems of existing image super-resolution algorithms,such as weak detail information recovery ability and unreasonable feature reuse,we propose a super-resolution reconstruction algorithm combining information distillation and double link sampling.First,multi-dimensional feature extraction is carried out through the multi-scale information distillation module so that the obtained feature information appears comprehensive and the representation ability of the network is enhanced.Second,multi-scale features are selectively extracted by distillation mechanism,and some distilled features are introduced into hierarchical attention mechanism for global reuse.During this process,not only the number of network parameters is reduced,but also the context information is enriched.Finally,features obtained under different paths are up sampled,and local and global features are combined to improve the ability of detail information recovery.Experimental results show the satisfactory image quality reconstructed by the proposed algorithm.For example,the average peak signal-to-noise ratio reaches 0.35 dB higher than that of the feather distillation interaction weighting network(FDIWN)at 4 times the magnification factor,the model number of parameters reaches 55%lower than that of the cascaded residual network(CARN),and the proposed algorithm outperforms the current mainstream lightweight algorithm.