Single image haze removal via attention-based transmission estimation and classification fusion network

摘要

With the rapid development of deep learning theory, many excellent convolutional neural network (CNN) based dehazing methods have been proposed for single image dehazing. However, the slow training con-vergence rate and haze residual are still two serious flaws of these existing dehazing networks. To tackle these issues, we propose a novel end-to-end CNN-based dehazing framework called attention-based transmission estimation and classification fusion network (ATECFN). The ATECFN framework consists of three submodules: attention-based transmission-airlight estimation network (ATAEN), multi-scale Auto-Encoders (MAE), and patch-based classification fusion network (PCFN). First, the transmission sim-ilarity, that is the similarity of neighboring pixels in transmission map, is introduced to significantly increase the capability of CNN to fit transmission map. Second, the ATAEN is exploited to estimate airlight map and transmission map and then they are used to obtain a rough dehazed result according to the atmospheric scattering model. Third, we present the MAE to further refine the rough result, where the multi-scale structure can effectively capture local details over a wide range of scales. Finally, PCFN, a new fusion strategy, is employed to integrate the two results generated by ATAEN and MAE, in which a probability map is derived from a binary classification network and viewed as the fusion coefficient map. Extensive experiments demonstrate that the proposed algorithm outperforms the state-of-the-art methods on both synthetic and real-world images, which can not only improve the training convergence rate but also remove the residual haze effectively.(c) 2021 Elsevier B.V. All rights reserved.