Multi-Images Restoration Method with a Mixed-Regularization Approach for Cognitive Informatics

Cognitive image processing is an important part of cognitive informatics. High quality images are crucial for cognitive image processing, especially in small object recognition and image segmentation. Multi-images restoration provides an alternative approach for these problems. For example, with image denoising and image deblurring, the raw images can be better provided to improve the result of cognitive image processing. The improvement of imaging device's sampling rate provides a clue to design a common approach for multi-images restoration. This paper concerns with a mixed-regularization approach for solving multi-images (MRMI) restoration problems. The MRMI algorithm generalizes the original total variation (TV) based algorithm by fusing multiple noisy images to maximize the useful information restored from the degraded images. The proposed approach combines l1 regularizer and TVp regularizer to restore a latent image, which operates on two different domains, i.e., pixel and gradient. This mixed-regularization method can simultaneously exploit the sparsity of natural signal. The resulting problem is solved by the adaptation of generalized accelerated proximal gradient (GAPG) method. The effectiveness of our approach is validated in the context of multi-images denoising, deblurring and inpainting. Compared with some iterative shrinkage-thresholding algorithms, the experimental results indicates that our approach can restore a better image.