Statistical Ratio Rank Ordered Differences Filter for SeaWiFS Impulse Noise Removal

Due to the 10-bit design of SEAWIFS instrument and the signal transmission, many SEAWIFS satellite images are expected to have lower digitization noise and corrupted impulse noise. In this paper, we first analyze the characteristics of impulse noise and propose a new rank-ordered filter based on the difference of sequence for mean/standard, which is named as Statistical Ratio Rank Ordered Differences (SRROD) filter. Second, we describe the impulse noise detection and removal algorithm in detail. Compared with the median filter and other existing filters, the SRROD filter could effectively remove impulse noises while preserving other valid pixels without, or only with minor, modification. Through adjusting the lower and upper threshold values, different filter performance could be achieved. We also discuss the blind parameters optimization for non-recursive implementation. Based on the assessment of the distribution map of performance estimator according to different lower and upper threshold pairs, a nearly optimal threshold could be obtained. Finally, some concluding remarks are also presented in this paper. Introduction During remote sensing image acquisition, transmission, and processing, due to the limitation of instrumental design, environmental and atmosphere conditions, channel transmission errors, signal encoding processing, and other reasons, images can be corrupted by some stochastic and randomly distributed black and white noises, which are usually called impulse noise, or salt and pepper noise. Salt and pepper noises severely degrade the image quality and limit quantitative assessment. Therefore, noise removal is very important for remote sensing imagery and other image processing, and it is also significant in improving the resultant effect of image segmentation, feature extraction, image recognition, and classification. At present, image filtering is the main method for removing noises from an image. The goal of impulse noise removal is to suppress the noise while preserving the integrity of edges and detail information. Conventional linear filters, such as mean filters, are not very effective for the removal of Statistical Ratio Rank Ordered Differences Filter for SeaWiFS Impulse Noise Removal Zhengjun Liu, Changyao Wang, Aixia Liu, and Xiangming Xiao salt and pepper noise; while some nonlinear filters, for example, the median filter (Turky, 1971; Lin and Willson, 1988) and order statistic (OS) filters (Kim, 1986, 1995; Bovik, 1983), can efficiently reduce most of the salt and pepper noise in the image with its edges and texture information degraded to varying degrees. Another limitation is the robustness of these algorithms; that is, the performance of the algorithms decreases significantly when the percentage of impulse noise in the image increases. To overcome this problem, some new algorithms have been recently proposed: the multistage median filters (Wendt et al., 1986; Coyle et al., 1988, 1989; Lin et al., 1990), center weighted median (CWM) filters (Hardie et al., 1993; Ko et al., 1991; Sun et al., 1992), general weighted median (WM) filters and weighted order statistic (WOS) filters (Yli-Harja et al., 1991), length adaptive median filter (Lin and Willson, 1988), decision-based median filter (Florencio and Schafer, 1994), stack filters (Coyle et al., 1988, 1989; Lin et al., 1990; Wendt et al., 1986), permutation filters (Barner et al., 1994), and rank-conditioned, rank-selection filters (Hardie, 1994). Most of these filters have demonstrated better performance than the median filter in the removal of impulse noise and detail preservation. However, because most of these approaches are typically implemented uniformly across an image, they also tend to modify pixels that are undisturbed by noises. In addition, some researchers introduced impulse noise filter using fuzzy logic techniques (Zhang and Wang, 1997; Wang and Zhang, 1998). These algorithms are based on fuzzy impulse detection and fuzzy noise cancellation techniques. Although there is some effective improvement, there are still difficulties to create the fuzzy rule like other fuzzy systems, especially when no training images are provided. In this paper, a novel, non-linear, adaptive algorithm is proposed for the removal of impulse noise from SEAWIFS images. The paper is organized as follows: the mathematical model of random-valued impulse noise is described; the detailed algorithm and image-processing technique based on Statistical Ratio Rank Ordered Differences (SRROD) filter is studied; the experimental results of applying our filter to SEAWIFS impulse noise removal is discussed; a non-iterative optimization of filter parameters is discussed; and finally, the conclusion. Impulse Noise Model and Algorithm Consideration Normally, impulse noise is a result of a random process in which the digital numbers (DNs) of the corrupted pixels are PHOTOGRAMMETRIC ENGINEER ING & REMOTE SENS ING J a n ua r y 2005 89 Z. Liu, C. Wang, and A. Liu are with the Laboratory of Remote Sensing Information Science, Institute of Remote Sensing Applications, CAS, Beijing 100101, China (zjliu@ casm.ac.cn). Z. Liu is also with the Institute for Photogrammetry and Remote Sensing, Chinese Academy of Surveying and Mapping, No.16 Beitaiping Road, Beijing 100039, China. X. Xiao is with the Complex System Research Center, Institute for the Study of Earth, Oceans, and Space, University of New Hampshire, Durham, NH 03824-3525. Photogrammetric Engineering & Remote Sensing Vol. 71, No. 1, January 2005, pp. 89–95. 0099-1112/05/7101–0089/$3.00/0 © 2005 American Society for Photogrammetry and Remote Sensing 02-083.qxd 2/20/04 11:27 PM Page 89