Single image dehazing based on the fusion of multi-branch and attention mechanism
Xiaohang Yu, Huikang Yu · 2022 International Conference on Big Data, Information and Computer Network (BDICN) · 2022
In recent years, convolutional neural networks have made significant progress in the field of image dehazing, but most methods still include the problems of incomplete dehazing and color distortion. An end-to-end dehazing network combining transfer learning and attention feature fusion mechanism is proposed to respond to these problems. First, we build a transfer learning sub-network by introducing a pre-training model. Secondly, two attention feature fusion molecular networks are used to assist the transfer learning network to obtain more accurate feature information. Finally, we use the tail convolution operation to achieve end-to-end dehazing processing, and the resulting graph is further obtained. The experimental results show that, compared with advanced algorithms, the dehazing effect of this algorithm is closer to clear images in both synthetic data sets and real data sets, and effectively solves the problem of color distortion.