DFF-Net: Single Image Dehazing with Attention-based Deep Feature Fusion Network

Abstract

We suggest end-to-end convolution neural network for recovering a haze-free image from contaminated image. The feature extraction module enables network to extract the features at various levels. The network provides additional flexibility in dealing with different types of information and focuses more on important information using mixed convolution attention mechanism. To improve the dehazing performance, multi-level features are fused and further refined using feature fusion block. The good kernel estimation can recover a sharp image. Moreover, DFF-Net has ability to capture sharp textural and semantic information, and recover high-quality haze-free image. Furthermore, semantic differences in deep features are measured by deep semantic loss. The experimental findings demonstrate that our suggested method exhibits superior performance compared to other haze-removal methods on both synthetic and real-world images.