Open AccessUDA‐Net: Densely attention network for underwater image enhancement
Author(s) -
Li Yang,
Chen Rong
Publication year - 2021
Publication title -
iet image processing
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.401
H-Index - 45
eISSN - 1751-9667
pISSN - 1751-9659
DOI - 10.1049/ipr2.12061
Subject(s) - underwater , computer science , artificial intelligence , pooling , benchmark (surveying) , convolutional neural network , feature (linguistics) , focus (optics) , channel (broadcasting) , image (mathematics) , computer vision , image quality , grid , pattern recognition (psychology) , mathematics , optics , geology , telecommunications , linguistics , oceanography , philosophy , physics , geometry , geodesy
Underwater imaging usually suffers from negative impacts due to the absorption and scattering effects in water. Underwater images thus have unfavourable visual quality to support the work in such environment. This paper addresses the problem of image improvement for single underwater image. The core idea lies in a new enhancement model based on deep learning architecture, in which a feature‐level attention model is developed. This model is a multi‐scale grid convolutional neural network that can facilitate fusing different types of information during representation learning. According to this information combination, a synergistic pooling mechanism is proposed to extract the channel‐wise attention maps to derive the locally weighted features. Therefore, this model can adaptively focus on the feature regions corresponding to degraded patches in one underwater image and improve these patches consistently. Comprehensive experiments are conducted on benchmark and natural underwater images, and it can be demonstrated that this model is effective.