Multi-level Feature Interaction and Efficient Non-Local Information Enhanced Channel Attention for image dehazing

Image dehazing is a challenging task in computer vision. Currently, most dehazing methods adopt the U-Net architecture that directly fuses the decoding layer with the corresponding scale encoding layer. These methods ignore the effective utilization of different encoding layer information and existing feature information dilute problems, resulting in suboptimal edge details and overall scene aspects of dehazed image restoration. In addition, Squeeze and Excitation (SE) channel attention is widely used in dehazing network. However, the two fully-connected layers of dimensionality reduction operation in SE will negatively affect the weight prediction of feature channels, thus reducing the performance of the dehazing network. To solve the above problems, we propose a Multi-level Feature Interaction and Non-local Information Enhanced Channel Attention (MFINEA) dehazing model. Specifically, a multi-level feature interaction module is proposed to enable the decoding layer to fuse shallow and deep feature information extracted from different encoding layers for better recovery of edge details and the overall scene. Furthermore, an efficient non-local information enhanced channel attention module is proposed to mine more effective feature channel information for the weight assignment of the feature maps. The experimental results on several challenging benchmark datasets show that our MFINEA outperforms the state-of-the-art dehazing methods.