Presentation Information
[A-2-10]Recursive Reservoir Concatenation for Random-Valued Impulse Noise Removal
◎Takahiro Kawano1, Michiharu Maeda1 (1. Graduate School of Engineering, Fukuoka Institute of Techology)
Keywords:
Reservoir computing,Random-valued impulse noise,Noise removal
This study presents a random-valued impulse noise removal by extending recursive reservoir concatenation (RRC), which has been originally developed for salt-and-pepper noise removal, with noise location estimation based on the median absolute deviation (MAD). Random-valued impulse noise is difficult to detect because pixel values are corrupted to arbitrary intensities, resulting in significant degradation of image quality. Our model estimates noise locations using MAD and trains the RRC model using only pixels estimated to be noise-free. Experimental evaluations were conducted using Peppers. The results demonstrate that our model effectively removes random-valued impulse noise while preserving image edges and fine structural details. Furthermore, we show that the MAD threshold significantly affects restoration performance and that an appropriate range of threshold values exists regardless of the image used.