We found a match
Your institution may have access to this item. Find your institution then sign in to continue.
- Title
U-SRN: Convolutional Neural network for single image super resolution.
- Authors
Dixit, Monika; Yadav, Ram Narayan
- Abstract
Single Image Super Resolution (SISR) aims to recover high-frequency details of an image from its low-resolution form. It is a highly challenging problem since low-resolution image is blur and contains noise. The application of SISR can be found in various fields, such as microscopic image analysis, medical imaging, security and surveillance imaging, biometric image identification, hyper spectral imaging and text image super-resolution. Convolutional Neural Networks (CNNs) are the most widely used technique for SISR. Most of the previous CNN-based SISR methods blindly increase the network depth to achieve good performance, leading to increased computational cost. They train feed-forward convolutional neural networks using a per-pixel loss between the output and ground-truth images. U-SRN model is proposed for single-image super-resolution, consisting of a contracting path to capture context and a symmetric expanding path that enables precise localization. Per-pixel loss is replaced with perceptual loss, which gives visually pleasing results. Extensive experiments show that the U-SRN model performs excellently and can perform multi-scale tasks. The model is trained on the DIV2K dataset. It is a benchmark dataset in the field of super-resolution, which consists of 1,000 high-quality images. Set 5, Set 14, Urban 100, and BSD 100 are the benchmark super-resolution datasets used for testing. For the upscaling factor of 4, the average gains on PSNR, SSIM and IFC achieved by U-SRN using Urban 100 dataset are 0.24 dB, 0.0037 and 0.025 higher than the next best approach. The results indicate that U-SRN performs better as compared to other state-of-art methods in terms of PSNR, SSIM and IFC on all datasets. The other performance analysis criterion, such as PI (4.12), FSIM (0.971), and SRQC (0.954), indicate the superiority of the model. The network has limited number of parameters and performs fast execution.
- Subjects
CONVOLUTIONAL neural networks; MICROSCOPY; SPECTRAL imaging; IMAGE analysis; BIOMETRIC identification; HUMAN fingerprints
- Publication
Multimedia Tools & Applications, 2024, Vol 83, Issue 16, p46875
- ISSN
1380-7501
- Publication type
Article
- DOI
10.1007/s11042-023-17379-2