Abstract:

The acquisition, transmission, and storage process of color images can corrupted by noise, which brings trouble to the subsequent image processing. The task of image denoising is of great significance. A model based on quaternion dictionary learning is proposed by using the method of constraining the real part of quaternion to zero to represent color images, so that the color information can be well preserved. The traditional sparse image models only regard color image pixels as the linear connection of vector or monochrome images, which ignore the correlation of RGB channels. The color image is represented as a pure quaternion, and the RGB channels of the color image are represented as the imaginary part of the quaternion matrix, which fits the image better. Compared with other advanced models, numerical experiments show that the constructed model can represent color images more accurately, and the model fitting error is relatively small in the process of processing color images, peak signal-to-noise ratio and visual effect are significantly improved, so it can better denoise the image.

Key words: quaternion, dictionary learning method, color image denoising