基于轻量级神经网络的单幅图像去雨滴模型

A single image raindrop removal model based on lightweight neural network

目前单幅图像去雨滴网络的特征图存在较高的相似性和冗余性,导致模型的参数量庞大,极大限制了其在实际应用中的部署。本文提出一种基于轻量级神经网络的单幅图像去雨滴模型。采用一种幻像特征生成残差块,用于解决网络中特征图的相似性和冗余性问题。设计了一种复合折叠式重用机制,有效改善了由于参数减少带来的模型性能下降。提出一种轻量级门控循环单元,用于强化折叠式去雨滴架构中的深度特征交互,进一步提高了模型的性能。实验结果表明:本文提出的轻量级去雨滴模型在性能持平或略高于目前3种算法的前提下,分别实现了模型参数量的18、37及51倍的压缩,较好解决了在实际应用中的部署问题。

At present,the feature maps in the single image raindrop removal network have high similarity and redundancy,resulting in a huge number of parameters of the existing models.It greatly limits the deployment in practical applications.To this end,this paper proposes a single image raindrop removal model based on lightweight neural network.The model employs a ghost feature generation residual block to solve the similarity and redundancy of feature maps in the network.A composite foldable reusing mechanism is designed to effectively improve the model’s performance degradation caused by parameter reduction.A lightweight gated recurrent unit is introduced to enhance the interaction of deep features in the folded raindrop removal architecture,which further improves performance of the model.The experimental results show that the proposed lightweight raindrop removal model respectively achieves 18-,37-,and 51-times the compression of parameters,whose performance is equal to or slightly higher than that of the current three state-of-the-art algorithms.It better solves the deployment problem in practical applications.