基于模块相似性的超分网络剪枝

Module Similarity-Based Pruning for Image Super-Resolution Network

  • 摘要: 该文针对单图像超分辨率网络(SISR)提出了一种简单的网络剪枝方法。该方法通过评估超分网络中各模块的相似性,用一种简单办法将相似度转换为各模块对网络的贡献程度,从而找到对超分网络相对不重要的模块进行网络剪枝,达到超分辨率网络压缩的目的。通过基于模块相似性的超分网络剪枝,原本参数量庞大的超分网络得到了压缩,参数量和运算量都大幅下降。实验表明,通过剪枝后的超分网络其参数量可以下降60%以上,同时精度下降不超过0.1%,对超分网络部署到低性能平台有着实际意义。

     

    Abstract: This paper proposes a network pruning method for single image super-resolution network (SISR). This method evaluates the similarity of each module in the super-resolution network and uses a simple method to convert the similarity into the contribution degree of each module to the network, and find the relatively unimportant modules of the network to perform network pruning. Through the method of network pruning for the super-resolution network based on the module similarity, the super-resolution network with a huge amount of parameters is compressed, and the number of parameters and the amount of calculation are greatly reduced. Experiments show that the parameters of the super-resolution network after pruning can be reduced by more than 60%, while the accuracy is not reduced by more than 0.1%, which has great practical significance for the deployment of the super-resolution network to a low-performance platform.

     

/

返回文章
返回