Abstract:

Dehazing algorithms based on training synthetic images cannot achieve satisfactory results on real image data sets. In view of these problems of unsatisfactory generalization ability, this study proposes a semi-supervised learning network based on residual attention mechanism for single image dehaze. The backbone network of the proposed model consists of an encoder and a decoder. Using stacked residual attention modules, the feature weights of different scales are adjusted to give more weight to important features. Local residuals choose to bypass hazy regions so that the model can focus on valid information. The training in this study is divided into two branches: supervised learning and unsupervised learning, which input synthetic data and real data respectively. Dark channel loss and total variational loss are used to constrain the unsupervised branches. The results show that the proposed algorithm obtains ideal results on both synthetic data sets and real data sets, and the average processing time of images is only 0.01 s, achieving a balance between dehazing effect and processing time.

Key words: image dehazing, encoding and decoding structure, semi-supervisory framework, attention mechanism, residual connection, SOS enhancement strategy, dark channel loss, SSIM loss