单幅图像去雾的多步融合自适应特征注意网络  

Self-Adaptation Feature Attention Network with Multi-Step Fusion for Single Image Dehazing

作　　者：张嘉伟 刘晓晨 赵东花 王晨光[4] 申冲[1,2] 唐军[1,2] 刘俊[1,2] ZHANG Jiawei;LIU Xiaochen;ZHAO Donghua;WANG Chenguang;SHEN Chong;TANG Jun;LIU Jun(Key Laboratory of Instrumentation Science & Dynamic Measurement(North University of China), Ministry of Education, Taiyuan 030051, China;Shanxi Province Key Laboratory of Quantum Sensing and Precision Measurement (201905D121001), North University of China, Taiyuan 030051, China;School of Instrument Science and Engineering, Southeast University, Nanjing 210096, China;School of Information and Communication Engineering, North University of China, Taiyuan 030051, China)

机构地区：[1]中北大学仪器科学与动态测试教育部重点实验室,山西太原030051 [2]中北大学量子传感与精密测量仪器山西省重点实验室(201905D121001),山西太原030051 [3]东南大学仪器科学与工程学院,江苏南京210096 [4]中北大学信息与通信工程学院,山西太原030051

出　　处：《测试技术学报》2022年第4期347-352,共6页Journal of Test and Measurement Technology

基　　金：国家自然科学基金创新研究群体资助项目(51821003);国家自然科学基金优秀青年基金资助项目(51922009);国家自然科学基金面上资助项目(61973281);山西省重点研发计划资助项目(202003D111003);山西省优秀青年培育资助项目(202103021222011);山西省1331工程项目。

摘　　要：雾天天气严重影响了人类户外活动的进行,学术界针对图像去雾的计算机视觉任务已经进行了广泛研究,但仍然面临着诸如真实雾图像去雾能力有限等严峻挑战.为此,提出了一种基于多步融合的端到端自适应特征注意网络.其中的自适应特征注意模块可以自适应扩展接收域,获取空间中的关键结构信息,提取更复杂的特征.此外,考虑到网络中获取的低层次和高层次特征之间缺乏连接,还完成了多步融合模块,该模块能使网络中不同层次的特征在图像恢复过程中有效互补.另外,通过减少网络参数,优化后的网络结构使得其所需的计算资源也大幅度减少.对于具有真实雾霾的Dense-Haze和NH-HAZE数据集,该方法得到了较高的峰值信噪比(PSNR)和结构相似度(SSIM),分别为16.23 dB,0.5213和21.38 dB,0.7144;同时,其实际视觉效果也优于其他所选先进技术.The existence of fog has seriously affected the human outdoor activities.For dehazing images,which is a common computer vision task,the academic circle has carried out extensive research.However,it still faces severe challenges such as the limited ability of real fog image to dehaze.Therefore,this study proposes an end-to-end adaptive feature attention network based on multi-step fusion.The adaptive feature attention module can adaptively expand the receptive field,obtain the key structure information in the space,and extract more complex features.In addition,considering the lack of connection between low-level and high-level features acquired in the network,the multi-step fusion module is also completed,which can make the features of different levels in the network effectively complement each other in the process of image restoration.Besides,by reducing network parameters,the computing resources required by the optimized network structure are greatly reduced.For dense haze and NH haze data sets with real haze,the research method obtains high peak signal-to-noise ratio(PSNR)of 16.23 dB and 21.38 dB and structural similarity(SSIM)of 0.5213 and 0.7144.The actual visual effects are better than other selected advanced technologies.

关 键 词：图像去雾 自适应特征注意 多步融合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]