Remote Sensing Image Denoising with Iterative Adaptive Wiener Filter

Abstract

Image denoising plays a significant role in the application of remote sensing images, since the noise not only deteriorates the visual quality, but also and more important, causes the performance drop of many computer vision algorithms, e.g., segmentation and object recognition. However, denoising is a quite challenging task, due to the complicated, nonlinear distribution of noises. In this paper, we propose an iterative adaptive Wiener filter for remote sensing image denoising, by exploring statistical characteristics of local similar patches. Given a noisy image, the proposed approach aims to pursuit a restored image, with sufficiently good quality. In the proposed method, firstly, a low-pass filter is applied to the observed noisy image. The resulted image is set as an initial version of the restored image, which is fed into the following iterative rounds and refined to progressively approximate to “noise-free” signal. In each round, we divide the image being processed into overlapping patches. Each one will be assigned into a group, by searching similar patches in its neighboring areas. Then the optimal Wiener filter model is estimated adaptively for each group, and performed on these involved patches. Since the sampled patches are overlapping, the resulted image is achieved by averaging on overlapped, filtered patches. After that, the resulted image will be processed in the same way in the next round. With the procedure repeated, the noises are gradually alleviated and the refined image is approached to “noise-free” one. Finally, the algorithm terminates when the image changes little of two neighboring iterations. The contribution of our paper lies in two aspects. First, we propose a novel Wiener filter strategy, which takes advantage of image self-similarity to estimate filter parameters adaptively. Second, iterative scheme can refine the results progressively, which significantly improve the image quality. Experimental results demonstrate that the proposed method outperforms state of the art methods and can significantly improve both the subjective and the objective quality of noisy remote sensing images.