Improving deep learning-based image super-resolution with residual learning and perceptual loss using SRGAN model

• Published in: Soft computing (Berlin, Germany) Vol. 27; no. 21; pp. 16041 - 16057
• Main Authors: Abbas, Rehman, Gu, Naijie
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.11.2023; Springer Nature B.V
• Subjects: Algorithms; Application of Soft Computing; Artificial Intelligence; Computational Intelligence; Control; Datasets; Deep learning; Engineering; Generative adversarial networks; Image enhancement; Image quality; Image resolution; Machine learning; Mathematical Logic and Foundations; Mechatronics; Medical imaging; Medical research; Methods; Neural networks; Perceptions; Performance evaluation; Robotics; Similarity; Structural integrity; Task complexity; Deep learning; Image quality; Perceptual loss; Structural similarity; Image super-resolution; Generative adversarial network; MSE loss; Residual learning; SSIM loss
• ISSN: 1432-7643; 1433-7479
• DOI: 10.1007/s00500-023-09126-4

This study introduces a new and inventive approach designed to address the complex challenges encountered in the domain of image super-resolution (SR) tasks based on deep learning. The super-resolution generative adversarial network (SRGAN) is an innovative architecture that integrates the concept of residual learning into the complex design of deep recursive neural networks. This integration is aimed at significantly improving the quality of generated images. To attain the intended improvement in image quality, a series of varied loss functions are used in progressive manner, incorporating the structural similarity index (SSIM) loss and the mean squared error (MSE) loss, which are grounded on the perceptual loss paradigm. The carefully crafted loss functions are designed to enhance and maintain the fundamental element of structural integrity in the generated images, which is a critical requirement for dependable image super-resolution. The proposed methodology encountered a thorough evaluation through a series of rigors assessments, using developed benchmark datasets that are frequently used in the relevant field. Through an evaluation of the model’s performance with respect to image quality and structural similarity, we have effectively determined its effectiveness and ability to enhance current efforts aimed at advancing the field of image super-resolution. The results obtained exhibit a high-level accuracy, with the rate of 99.04%, and a less loss value of 3.19%. by achieving SSIM score of 0.97, PSNR score of 34.6, and MSE score of 0.012.