基于递归金字塔的多任务轻量化图像超分辨率算法

Multi-task Lightweight Image Super-resolution Algorithm Based on Recursive Pyramid

随着深度学习在超分辨率领域的广泛研究,为提高重建精度,网络结构呈现参数量越来越大的特点,且单个网络只能解决单个放大任务。针对这些问题,提出了基于递归金字塔的多任务轻量化图像超分辨率算法。算法中设计的递归金字塔模块以递归网络为基础,共享特征映射主体的权重使用独立的重建模块完成多个×2的超分子任务。特征映射主体以局部自适应融合模块堆叠而成,采用密集连接和特征自适应融合的方式,实现了特征的高效提取和参数的轻量化。本文方法在Set5、Set14、B100和Urban100上与现有算法进行比较,实验结果表明,所提方法在视觉上重建效果更好,参数量和浮点运算量更少。

With the extensive research of deep learning in the field of super-resolution,in order to improve the reconstruction accuracy,the network structure is characterized by an increasing number of parameters,and a single network can only solve a single amplification task.To solve these problems,a recursive pyramid based multi-task lightweight image super resolution algorithm is proposed.The recursive pyramid module designed in the algorithm is based on the recursive network,and the weight of the shared feature mapping subject is completed by using the independent reconstruction module to complete several×2 supramolecular tasks.The main body of feature mapping is composed of local adaptive fusion modules,which realizes efficient feature extraction and lightweight parameters by means of dense connection and adaptive feature fusion.The proposed method is compared with the existing algorithms on SET5,SET14,B100 and URBAN100,and the experimental results show that the proposed method has better visual reconstruction effect and less parameters and floating point computation.