雨天场景下单目图像深度估计与清晰化算法  

Depth Estimation and Clarification Method for Monocular Images in Rainy Scenes

作　　者：张家豪[1] 张娟[1] 郎晓奇 ZHANG Jia-hao;ZHANG Juan;LANG Xiao-qi(School of Electronic and Electrical Engineering,Shanghai University of Engineering Science,Shanghai 201620,China)

机构地区：[1]上海工程技术大学电子电气工程学院,上海201620

出　　处：《小型微型计算机系统》2023年第11期2584-2590,共7页Journal of Chinese Computer Systems

基　　金：地方院校能力建设项目(21010501500)资助;上海市科委科技创新行动计划(22S31903700)资助。

摘　　要：场景深度估计在三维重建、自动驾驶等应用中极为重要,目前深度估计缺乏在恶劣天气条件下的应用研究,实际场景表现不佳.本文针对雨天场景进行研究提出了一种联合的场景深度估计和图像去雨算法.其中深度估计网络以Transformer结构作为编码器和解码器主要模块,首先通过对图像块进行重排和线性投影生成嵌入块,减少了下采样的特征损失;接着利用多头自注意力机制在不同尺度提取特征并与解码器通过跳跃连接对局部和整体深度特征进行学习,提高了全局和长距离上下文信息的利用率,在RainCityscapes数据集上的场景深度估计质量优于现有算法.此外本文还将深度图结果作为先验信息,通过深度信息引导全局残差特征融合去雨网络得到无雨图像,在多个公开数据集上相比现有去雨算法的结构相似度(SSIM)和峰值信噪比(PSNR)均有提高.Scene depth estimation is extremely important in applications such as 3D reconstruction and autonomous driving.Currently published estimates of the application of research in bad weather,the actual performance is poor.In this paper,a joint scene depth estimation and image deraining algorithm is proposed for rainy scenes.Among them,the depth estimation network uses Transformer structure as the main module of encoder and decoder,and firstly,the embedding block is generated by rearranging and linearly projecting the image blocks,which reduces the feature loss of down-sampling,then features are extracted at different scales using multi-head selfattention mechanism and learned with decoder through jump connection for local and overall depth features,which improves the global and long-range contextual information utilization rate,and the quality of scene depth estimation on the RainCityscapes dataset is better than existing algorithms.In addition,this paper also uses the depth map results as a priori information to guide the global residual feature fusion deraining network to obtain rain-free images by depth information,which improves the structural similarity(SsIM)and peak signal-to-noise ratio(PSNR)compared with existing deraining algorithms on several public datasets.

关 键 词：深度估计 图像去雨 多尺度特征融合 深度引导 全局特征 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]