Transpose convolution based model for super-resolution image reconstruction

doi:10.1007/s10489-022-03745-4

Transpose convolution based model for super-resolution image reconstruction

2022-08-20

15

PRE

AI

DOI：10.1007/s10489-022-03745-4

Original

Original request for help

Save

Abstract

En 中文

Single image resolution is a noticeably challenging issue that targets to acquire a high-resolution output out of one of its low-resolution variants. Many existing approaches for single-image resolutions are based on the direct solving details by using pre-defined up sampling operators. Therefore, it is challenging for the reconstruction process when the image has a larger upsampling factor. Recently, convolution neural networks (CNNs) made easy progress on super-resolution (SR) image with good results. However, the majority of methods are based on pre-defined up sampling, which uses the bicubic interpolation technique for upscaling the low-resolution (LR) image and employs feature maps to reconstruct the final high-resolution (HR) image. This leads to visual artifacts in reconstructed images and can be difficult to train such a model with a larger network. Therefore, we remove the proposed transposed convolution layer method with a novel architecture and avoid the usage of pre-defined up sampling operators. We purpose an efficient method for the usage of transposed convolution with a new architecture design and use a recurrent residual block for mapping extraction in a step-by-step manner. Finally, we generate the desired super-resolution image with low complexity and fewer parameters. Experiments and state-of-art results show better performance than existing models.

Keywords:

Super-resolution

CNN

Deep learning

Recurrent residual learning

Multi-level output

Transpose convolution based model for super-resolution image reconstruction

Abstract

Journal

Organization