Recursive Gated Convolution Based Single Image Dehazing Network

The conventional prior knowledge based dehazing algorithm can recover the haze image, yet its dehazing performance would get worse even unattainable in the case that the required prior knowledge is unsatisfied. To this end, the deep learning (DL) based dehazing networks have becoming a promising technique in such a haze removal topic. However, most of the existing DL-based dehazing networks only employ the common convolution to establish the model, which does not involve the high-order spatial interaction information. Therefore, we integrate feature extraction module, having high order inter-active feature extraction block, and multi-scale feature attention module to propose a recursive gated convolution based single image dehazing network. Specifically, the high order spatial interaction, which is beneficial to improving the modeling ability of dehazing model, is conducted through designing the gated convolution and recursive. Meanwhile, we construct the multi-scale feature attention module by introducing the global and local channel attention mechanism, effectively realizing the recovery and enhancement of image structure information and edge detail information. Finally, we design a feature reconstruction module to restore the haze removal image. The experimental results verify that, compared with benchmarks, the proposed dehazing network not only achieves a better objective evaluation score, but also reconstructs a better subjective dehazing image.