Unsupervised single image dehazing network

Abstract

In image dehazing algorithms, most algorithms use supervised learning method for image dehazing. Because the ideal data set is difficult to collect, many supervised learning methods use composite data set, but the effect of the model trained in composite data set is often not ideal when applied to the real-world images. To solve this problem, a single image dehazing network has been proposed for the purpose of unsupervised single image dehazing network. The contributing factors include: (1) solving the problem of hard acquisition of ideal data sets, (2) proposing unsupervised loss functions, (3) introducing a method for embedding a model unsupervised neural network. We create unsupervised loss functions according to existing knowledge: the bright channel loss, the dark channel loss and the dark channel energy loss, we propose unsupervised loss functions that guarantee image dehazing algorithms. Since there are many experiments on synthetic and real-world images, the proposed method can improve the details, structure and texture of the image by better modifying the three advanced dehazing methods.