TY - JOUR
T1 - Distortion detection and removal integrated method for image restoration
AU - Wang, Yuhang
AU - Li, Hai
AU - Hou, Shujuan
N1 - Publisher Copyright:
© 2022
PY - 2022/7
Y1 - 2022/7
N2 - Image restoration has been the focus of research in image processing, and current methods mainly target specific single distortion or hybrid distortion with known distortion types. However, real-world images generally affected hybrid distortions of unknown quantity and type, there is no restoration method applicable to such complex hybrid distortion. Therefore, we propose a distortion detection and removal integrated method. Firstly, the distortion detection module is designed based on the idea of multi-label classification, which can detect the type and level of distortion in the distorted image. The type and level of distortion can be used as a piece of prior information to control the subsequent image restoration process. Then, the image restoration module is designed based on controllable residual connection, which enables the network to restore images with different types and levels of distortion, and the attention parallel convolution block is designed by using a parallel convolution layer and coordinated attention mechanism to enhance the feature extraction ability of the network and improve the quality of the restored image. The experimental results show that the proposed method achieves superior performance.
AB - Image restoration has been the focus of research in image processing, and current methods mainly target specific single distortion or hybrid distortion with known distortion types. However, real-world images generally affected hybrid distortions of unknown quantity and type, there is no restoration method applicable to such complex hybrid distortion. Therefore, we propose a distortion detection and removal integrated method. Firstly, the distortion detection module is designed based on the idea of multi-label classification, which can detect the type and level of distortion in the distorted image. The type and level of distortion can be used as a piece of prior information to control the subsequent image restoration process. Then, the image restoration module is designed based on controllable residual connection, which enables the network to restore images with different types and levels of distortion, and the attention parallel convolution block is designed by using a parallel convolution layer and coordinated attention mechanism to enhance the feature extraction ability of the network and improve the quality of the restored image. The experimental results show that the proposed method achieves superior performance.
KW - Attention mechanism
KW - Controllable residual connection
KW - Hybrid distortion
KW - Image restoration
UR - http://www.scopus.com/inward/record.url?scp=85128220085&partnerID=8YFLogxK
U2 - 10.1016/j.dsp.2022.103528
DO - 10.1016/j.dsp.2022.103528
M3 - Article
AN - SCOPUS:85128220085
SN - 1051-2004
VL - 127
JO - Digital Signal Processing: A Review Journal
JF - Digital Signal Processing: A Review Journal
M1 - 103528
ER -