Denoising of multichannel images with references

In this paper, we study a problem of filtering noisy component image of a multichannel image, under assumption that the multichannel data contain almost noise-free component image(s) highly correlated with the noisy one. Our proposed denoising approach is based on three-dimensional (3D) representation of the noisy and reference images. One dimensional discrete cosine transform (DCT) is applied to decorrelate images and then the obtained data are processed by the BM3D filter in the component-wise manner. Our approach has another option where the modified BM3D filter is applied. Performances of these methods are analyzed for ten test images, several values of noise variance and different quality metrics. It is demonstrated that performance depends on a choice of the reference images and the way they are preprocessed. In the case of proper pre-processing, improvements of the metrics PSNR and PSNRHVS- M can reach up to 3-7 dB compared to the component-wise BM3D filtering of the noisy component image. Examples of processing real-life hyperspectral images are presented with the recommendations on how to choose and pre-process reference images. High efficiency and relative simplicity of the proposed approach is demonstrated.