A Modern Approach to High Dynamic Range Image Processing with Machine Learning Architectures

In this paper, we analyze modern approaches and methods of neural networks employment in the tasks of capturing and demonstrating a wide dynamic sound stage (HDR). We highlight the essence of the problem and its relevance to the modern image processing applications. We identify the limitations of the conventional HDR image processing and discuss how the use of neural networks allows to address these limitations. We overview the use of various types of Machine Learning architectures in HDR image processing, such as convolutional neural networks, generative adversarial networks, and recurrent neural networks. In addition, we discuss unsupervised learning-based methods also considered by modern research. Based on our review, we develop recommendations of employing various HDR image processing methods in distinct real-world scenarios.