Deep-Aware Network for Removing Single Haze

Some dehazing methods based on traditional statistical theory are prone to suffering from the higher complexity and resulting in some lower degeneration phenomenon such as a few halo with discontinuities and low contrast. To tackle these issues, we present an effective and efficient end-to-end method based on deep-aware channel information for single image removal haze. We compute an effective feature map by integrating the deep-aware channel attention mechanism with multi-scale residual connection to build a feature aware attention block, and then more these blocks are combined to form the final feature block for directly learning the statistical information of some haze-free images. In some pubic synthetic datasets and some real-world haze images, extensive experimental results show that the proposed model is not only superior than some state-of-the-art models including some traditional dehazing methods and methods based on convolutional neural network, but also achieves better image quality than other models such as high contrast, better details, and vivid color. Meanwhile, we argue that our model is a lightweight and low computation cost network.