Denoising Color Images Based on Local Orientation Estimation and CNN Classifier

Abstract

A structure-adaptive vector filter for removal of impulse noise from color images is presented. The proposed method is based on local orientation estimation. A color image is represented in quaternion form, and then, quaternion Fourier transform is used to compute the orientation of the pattern in a local neighborhood. Since the computation in quaternion frequency domain is extremely time-consuming, we prove a theorem that the integral of the product of frequency variables and the magnitude of quaternion frequency signals can be computed directly in spatial domain, which results that the color orientation detection problem can be solved in spatial domain. Based on the local orientation and orientation strength, the size, shape, and orientation of the support window of vector median filter (VMF) are adaptively determined, leading to an effective structure-adaptive VMF. Unlike the classical VMF restricting the output to the existing color samples, this paper computes the output of VMF over the entire 3D data space, which boosts the filtering performance effectively. To further improve denoising effect, a deep convolutional neural network is employed to detect impulse noise in color images and integrated into the proposed denoising framework. The experimental results exhibit the effectiveness of the proposed denoiser by showing significant performance improvements both in noise suppression and in detail preservation, compared to other color image denoising methods.