A Novel Feature Attention Network for Single Image Dehazing
Yue Yang, Fengdan Lao, Chengtao Zhao, Weichao Yi · 2024
Image dehazing presents a formidable challenge within the domain of fundamental visual processing tasks since haze severely degrades the image quality and hampers the practical application. Therefore, tackling this issue is deemed essential and imperative. A novel feature attention network is presented in a seamless manner, designated as FANet tailored specifically for enhancing single image dehazing, comprises three distinct stages, comprising encoder stage, share stage, as well as decoder stage. With the paramount importance of contextual information in the dehazing process, a new block has been introduced by us, referred to as the contextualized dilated attention residual block (CDARB), serving as the central element in feature extraction. The block utilizes varying scales of dilated convolutions to expand the receptive field and a feature attention module that dynamically adjusts feature responses for both channel and spatial data. Consequently, we introduce a multifaceted loss formulation designed to enhance the performance of our proposed network. Comprehensive experimental assessments have validated that FANet consistently matches or surpasses the efficacy of preceding top-tier methodologies, as demonstrated by a broad range of quantitative and qualitative metrics.