J Shanghai Jiaotong Univ Sci ›› 2021, Vol. 26 ›› Issue (4): 543-553.doi: 10.1007/s12204-020-2203-2

• • 上一篇    下一篇

Distribution-Transformed Network for Impulse Noise Removal

LI Guanyu (李冠玉), ZHANG Fengqin (张凤芹), LIU Qiegen * (刘且根)    

  1. (Department of Electronic Information Engineering, Nanchang University, Nanchang 330031, China)
  • 出版日期:2021-08-28 发布日期:2021-06-06
  • 通讯作者: LIU Qiegen * (刘且根) E-mail: liuqiegen@ncu.edu.cn

Distribution-Transformed Network for Impulse Noise Removal

LI Guanyu (李冠玉), ZHANG Fengqin (张凤芹), LIU Qiegen * (刘且根)    

  1. (Department of Electronic Information Engineering, Nanchang University, Nanchang 330031, China)
  • Online:2021-08-28 Published:2021-06-06
  • Contact: LIU Qiegen * (刘且根) E-mail: liuqiegen@ncu.edu.cn

摘要: This work aims to explore the restoration of images corrupted by impulse noise via distributiontransformed network (DTN), which utilizes convolutional neural network to learn pixel-distribution features from noisy images. Compared with the traditional median-based algorithms, it avoids the complicated pre-processing procedure and directly tackles the original image. Additionally, different from the traditional methods utilizing the spatial neighbor information around the pixels or patches and optimizing in an iterative manner, this work turns to capture the pixel-level distribution information by means of wide and transformed network learning. DTN fits the distribution at pixel-level with larger receptions and more channels. Furthermore, DTN utilities a residual block without batch normalization layer to generate a good estimate. In terms of edge preservation and noise suppression, the proposed DTN consistently achieves significantly superior performance than current state-of-the-art methods, particularly at extreme noise densities.

关键词: impulse noise removal , deep learning , convolutional neural network , distribution transformation

Abstract: This work aims to explore the restoration of images corrupted by impulse noise via distributiontransformed network (DTN), which utilizes convolutional neural network to learn pixel-distribution features from noisy images. Compared with the traditional median-based algorithms, it avoids the complicated pre-processing procedure and directly tackles the original image. Additionally, different from the traditional methods utilizing the spatial neighbor information around the pixels or patches and optimizing in an iterative manner, this work turns to capture the pixel-level distribution information by means of wide and transformed network learning. DTN fits the distribution at pixel-level with larger receptions and more channels. Furthermore, DTN utilities a residual block without batch normalization layer to generate a good estimate. In terms of edge preservation and noise suppression, the proposed DTN consistently achieves significantly superior performance than current state-of-the-art methods, particularly at extreme noise densities.

Key words: impulse noise removal , deep learning , convolutional neural network , distribution transformation

中图分类号: