摘要
CNN网络深度的增加,导致计算成本急剧提升,且深层网络不能充分利用浅层特征.针对这个问题,提出了注意力机制引导下的特征增强网络(AGFENet),主要包括扩展卷积块(DVB)、特征增强块(FEB)和注意块(AB).DVB采用扩张卷积来扩大卷积核的感受野,有效降低网络深度,权衡性能和效率.FEB使浅层特征信息更多地流向深层网络,提高网络特征表达能力.AB引导FEB与DVB进行注意力机制处理,对厚薄雾区域采取不同的计算权重,充分提取图像厚雾的信息.实验表明,AGFENet的网络层数有效降低,针对合成和真实有雾图像集,定量和定性评价表现均较好.
With the increase of the depth of CNN network,the cost of computing increases sharply,and the deep network can not make full use of the shallow features.To solve this problem,It proposes a feature enhanced attention guided single image defogging(AGFENet),which mainly includes extended convolution block(DVB),feature enhancement block(FEB)and attention block(AB).DVB uses expanded convolution to expand the receptive field of convolution core,effectively reduce the network depth,and trade-off performance and efficiency.FEB makes the shallow feature information flow more to the deep network,and improves the feature expression ability of the model.AB guides FEB and DVB to process the attention mechanism,and takes different weights to calculate the thick and thin fog area,so as to fully extract the thick fog information of the image.The experimental results show that the number of network layers of AGFENet is effectively reduced,and the performance of quantitative and qualitative evaluation is good for synthetic and real foggy image sets.
作者
何胜敏
陈志翔
HE Shengmin;CHEN Zhixiang(Computer College,Minnan Normal University,Zhangzhou,Fujian 363000,China;Key Laboratory of Data Science and Intelligence Application,Minnan Normal University,Zhangzhou,Fujian 363000,China;College of Physics and Information Engineering,Minnan Normal University,Zhangzhou,Fujian 363000,China)
出处
《闽南师范大学学报（自然科学版）》
2021年第3期55-61,共7页
Journal of Minnan Normal University：Natural Science
基金
漳州市自然科学基金(ZZ2020J33)
闽南师范大学研究生教改课题(MSYJG8)。
