Universal image enhancement algorithm using deep neural neworks

Abstract

The article considers the problem of image quality improvement under the influence of various types of noise and distortion. A comparative analysis of standard algorithms and neural network algorithms for improving image quality based on deep neural networks has been carried out. The neural network algorithms are considered to be a universal mean for improving image quality. This article also offers the investigation of the influence of hyperparameters of deep neural networks on the quality of reconstructed images. The first part deals with theoretical issues of the image restoration problem which is regarded as a solution to a complex, unstable optimization task. Various augmentation techniques based on the application of methods of forced image noise and artificial defect generation are also considered. Finally, we introduce a new approach for data augmentation by partial image stylization. In addition, we use various metrics for assessing the quality of restored images, determine their disadvantages and limits of their applicability. In the second part of the article, the well-known classical algorithms are compared with the proposed neural network algorithms, and the dependence of image recovery quality from noise level is investigated. We evaluate the effectiveness of the proposed approaches for image improvement, consider additive and pulse types of noise, which can be found on medical images, as well as on images obtained using optical coherence tomography. The possibility of algorithms working in real time is also investigated. We propose the method of comparing the quality of images through the segmentation task by using the Unet network. The research procedures have enabled us to conclude that neural networks are not inferior to classical algorithms in improving image quality, and in some cases even outperform them.