Dependence comparison of the effectiveness of neural networks to improve image resolution on format and size
Abstract
Dependence comparison of the effectiveness of neural networks to improve image resolution on format and size
Incoming article date: 28.02.2024Roads have a huge impact on the life of a modern person. One of the key characteristics of the roadway is its quality. There are many systems for assessing the quality of the road surface. Such technologies work better with high-resolution images (HRI), because it is easier to identify any features on them. There are a sufficient number of ways to improve the resolution of photos, including neural networks. However, each neural network has certain characteristics. For example, for some neural networks, it is quite problematic to work with photos of a large initial size. To understand how effective a particular neural network is, a comparative analysis is needed. In this study, the average time to obtain the HRI is taken as the main indicator of effectiveness. EDSR, ESPCN, ESRGAN, FSRCNN and LapSRN were selected as neural networks, each of which increases the width and height of the image by 4 times (the number of pixels increases by 16 times). The source material is 5 photos of 5 different sizes (141x141, 200x200, 245x245, 283x283, 316x316) in png, jpg and bmp formats. ESPCN demonstrates the best performance indicators according to the proposed methodology, the FSRCNN neural network also has good results. Therefore, they are more preferable for solving the problem of improving image resolution.
Keywords: comparison, dependence, effectiveness, neural network, neuronet, resolution improvement, image, photo, format, size, road surface