首页|融合多尺度特征信息的图像雨滴去除方法

融合多尺度特征信息的图像雨滴去除方法

Image raindrop removal method fused with multi-scale feature information

扫码查看
原文链接
  • NETL
  • NSTL
  • 万方数据
针对雨滴使雨天图像背景特征模糊失真的问题,提出一种融合多尺度特征信息的图像雨滴去除算法.首先,搭建了一个编码-解码神经网络来学习图像特征映射,考虑到雨滴的物理形状特征,采用雨滴形状驱动注意力模块来捕捉雨滴位置.然后,引入空间与通道协调注意力机制,加强图像重要空间和通道特征权重.接着,利用空洞卷积、非对称卷积和金字塔结构设计了新型空洞空间卷积池化金字塔模块,以捕获图像的多尺度特征.最后,在同尺度的编码-解码卷积层间加入跳跃连接,将特征信息馈送到网络深处,达到去除图像中雨滴的目的.实验结果表明:本文算法在公开数据集Qian上的PSNR达到30.75,SSIM达到0.925 7;在自制雨天数据集上也可以有效去除图像中的雨滴.
Aiming at the problem that the existence of raindrops makes the background features of rain images blurred and distorted,an image raindrop removal algorithm that integrates multi-scale feature information is proposed.First,an encoder-decoder neural network is built to learn image feature mapping.Considering the physical shape characteristics of raindrops,the raindrop shape is used to drive the attention module to capture the raindrop position.Then,a spatial and channel coordinated attention mechanism is introduced to strengthen image important spatial and channel feature weights.Then,a novel atrous spatial convolution pooling pyramid module is designed using atrous convolution,asymmetric convolution and pyramid structure to capture multi-scale features of images.Finally,skip connections are added between the encoding-decoding convolutional layers of the same scale,and the feature information is fed to the depth of the network to achieve the purpose of removing raindrops in the image.The experimental results show that the algorithm in this paper has a PSNR value of 30.75 and an SSIM value of 0.925 7 on the public dataset Qian.The raindrops in the image can also be effectively removed on the self-made rainy day dataset.

image derainingdeep learningdilated convolutionspatial and channel coordinated attention mechanismencoding-decoding structure

崔明义、冯治国、代建琴、赵雪峰

展开 >

贵州大学机械工程学院,贵州贵阳 550025

图像去雨 深度学习 空洞卷积 空间与通道协调注意力机制 编码-解码结构

贵州省科技重大专项贵州省交通运输厅科技项目贵州省交通运输厅科技项目

ZNWLQC[2019]30122019-312-0202021-322-021

2024

10.19304/J.ISSN1000-7180.2023.0238

微电子学与计算机
中国航天科技集团公司第九研究院第七七一研究所

微电子学与计算机

CSTPCD
影响因子：0.431
ISSN：1000-7180
年,卷(期)：2024.41(4)
  • 1
  • 16