Single Image Dehazing Using Frequency Attention

Image degradation is a negative impact on computer vision tasks, and single image dehazing methods based on data-driven have witnessed the continuously growing and achieved great success. However, most data-driven methods are based on convolution neural network (CNN) which can be considered as be an implicit modeling of frequency domain. While discrete cosine transform (DCT) can be used in the CNN to model the features in frequency domain explicitly. Therefore, we propose an end-to-end image dehazing network with frequency attention (FA) based on DCT. Then, we select the top-K low-frequency components as the output of DCT layer, so the FA module can extract information from different frequency components. The experiments on the benchmark RESIDE demonstrate that our methods achieve better results than the previous state-of-art methods.