Professor Jin Qiyu was invited to give a lecture

发布时间: 2023-03-20 00:53:20

Topic：The theory and extension of the non-local means algorithm

Speaker: Jin Qiyu

Time: Wednesday, April 5th, 2023, 10:00-11:00 AM (Beijing Time)

Tencent Meeting ID: 669-2471-6718

Introduction:

The non-local means algorithm has been widely used in various fields of artificial intelligence due to its excellent denoising effect and combinability. This report discusses the convergence theory of the non-local means algorithm, and theoretically solves the problem of window parameter selection for the non-local means algorithm, greatly improving the algorithm's performance. Furthermore, the smooth kernel of the non-local means algorithm is studied, and the adaptive algorithm NLM-L2 and extended non-local means algorithm are obtained by minimizing the mean square error of the smooth kernel parameters. In order to generalize the algorithm, Gaussian noise is weakened into independent and identically distributed noise with zero mean (including Laplace noise and Poisson noise after variance-stabilizing transformation). In addition, by minimizing the absolute error, the corresponding smooth kernel parameter adaptive algorithm NLM-L1 is also obtained, which is proved to be more difficult than NLM-L2 in theory, but NLM-L1 has stronger robustness. Based on the theoretical research, a new algorithm combining the non-local means algorithm with low-rank matrix decomposition has been developed, which has excellent denoising effect in traditional algorithms and excellent effect of matrix completion with random point accuracy.

Organizers: South China Normal University, School of Mathematical Sciences, Machine Learning and Optimization Computing Laboratory Co-organizer

Guangdong Young Scientists Association