SINGLE IMAGE DEHAZING USING PHYSICS-INFORMED CONVOLUTIONAL AUTOENCODER

Abstract

Background. Generally, haze can be considered to be one of the most fundamental phenomena causing image visibility degradation. Numerous haze removal approaches have been proposed and most of them have achieved significant progress. In this paper the problem of dehazing images using deep neural network technologies is raised. Materials and methods. Our method is based on a physics-informed convolutional autoencoder. To achieve the best results in image dehazing, the black channel prior was integrated into the loss function. The model was trained on synthetically obtained pairs of clean and hazy indoor and outdoor images. The efficiency of the developed method was compared with AOD-net, CAP and Dehaze-Net. Results. It is established that the algorithm developed as part of the study is able to dehaze dense images. On the average, the method copes well with its task and is not inferior to large-scale neural networks. Conclusions. It was decided to improve the parameters of the model, as well as make small changes to the learning process to eliminate the problem of image lightening having large celestial areas.

Keywords

• artificial intelligence
• convolutional neural network
• autoencoder
• physicsinformed neural network
• image dehazing
• dark channel prior