Deeper super-resolution generative adversarial network with gradient penalty for sonar image enhancement

Shen, Pengyang; Zhang, Liguo; Wang, Minghao; Yin, Guisheng

doi:10.1007/s11042-021-10888-y

Cited by 12 publications

(1 citation statement)

References 33 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…From the experimental analysis, the outcomes of the study stated that, the suggested model obtained superior outcomes in inter-scanner and inter-protocol translation than other prevailing methods. [18] suggested a Deep super resolution GAN (DGP SRGAN) with gradient penalty for producing high resolution sonar images. Further, the major contribution of this study was addition of gradient penalty to loss function, which has resulted in faster and more stable training network.…”

Section: Introductionmentioning

confidence: 99%

Medical Image to Image Translation using Multi Label Activated Gradients for General Adversial Networks ((MLAG GAN))

Raja

2022

Preprint

View full text Add to dashboard Cite

Multi modal MRI provides complementary and clinically relevant information from the image to understand the condition of the tissue and to characterize various complex diseases. But imaging artifacts influence the determination of relevant inforamtion from brain metastatis which is difficult to obtain adequate number of modalities from same study subject because of the under optimized study plan. However quantitative analysis becomes mandatory for in-depth understanding of the disease. Existing works could not exploit and maintain texture details from the scanners. In this research work, it is designed a multi label activated gradients for GAN (MLAG GAN) to enrich the corresponding reconstruction images with huge information. The detailed study trained the developed model by permitting the gradient flow from multiple gradients to a single generator at multiple labels thereby addressing the prevailing limitaiton of GAN. This system could exploit multi label neural transfer that enable to grasp more number of semantic and lesion related priors from the reference images. It is also validated the proposed system on Flair T1 and Flair T2 from Brats 18 dataset that depicts superior performance on the quality of image generation when compared to the state of art methods. The multi label GAN regenerates more high resolution structure and texture information and the wide range of qualitative and quantitative comparative experiments with the state of art methods proves the effectiveness of the proposed system in terms of L1, MSE (Mean square error), PSNR (Peak to signal ratio),SSIM (Structure similarity index measure), and Inception Score( IS).

show abstract

Section: Introductionmentioning

confidence: 99%