A Cluster-Based Adaptive Switching Median Filter Yunfan Wang School of Instrument Science and Engineering Southeast university Nanjing, China Yunfanwang2003@gmail.com Zhu Zhu School of Instrument Science and Engineering Southeast university Nanjing, China edileon1@gmail.com Lei Miao School of Mechanical Engineering Southeast university Nanjing, China leoseu@163.com Xiaoguo Zhang * School of Instrument Science and Engineering Southeast university Nanjing, China zxg519@sina.com Xueyin Wan School of Instrument Science and Engineering Southeast university Nanjing, China xueyin_wan@sina.com Qing Wang School of Instrument Science and Engineering Southeast university Nanjing, China w3398a@263.net Abstract—This paper presents a cluster-based adaptive weight switching median filter. Clustering analysis and a linear function is combined to capture local image statistics. In term of the local information, an iteration function is constructed to subtract impulses from corrupted image and thus noise detector is defined. After the noisy pixels are identified, in order to keep image details as intact as possible, a cluster- based adaptive weighted median filter is proposed to estimate those noise candidates’ values. Simulation results show that the proposed method provides better performance in term of PSNR and MAE than many existing random-valued impulse noise filtering techniques. Keywords- Clustering; impulses; image details I. INTRODUCTION One of the most frequent problems during image acquisition and transmission is contamination of images by impulses noise due to noisy sensor or channel transmission errors [1]. The quality of an image affects the performance of image- processing techniques, such as edge detection, pattern recognition, and image segmentation. Therefore noise reduction and image restoration are essential in image- processing field. Generally, there are mainly two types of impulse noise models: the fixed-valued impulse noise and the random-valued impulse noise. An important characteristic of this type of noise is nonlinear, that means only parts of the pixels are corrupted and the rest are noise free. Comparing with random-valued noise, the fixed-valued noise is simpler and easier to restore for its gray-level value either takes minimal or maximal [2, 3], while the gray-level value of random-valued impulse noise is uniformly- distributed between minimal and maximal. In this paper, we mainly focus on processing random-valued impulse noise. Due to the extremely nonlinear nature of the impulse noise, a number of nonlinear approaches have been proposed for removing it. The standard median (SM) filter 4] is one of efficient nonlinear techniques widely used. However, since each pixel in the image is replaced by the median value in its neighborhood, SM filter is prone to damaging important details such as thin lines and sharp corners especially when the image is high corrupted. To this end, many improved median filter techniques have been proposed. Among them weighed-based median filters and the switching-based median filters are two typical solutions. The weighted median filters [5, 6] can perform different amounts of smoothing on different pixels by assigning different weights to their neighborhood pixels and thus they could effectively preserve fine image details while suppressing impulses. In addition, in order to increase details and sharpness preservation and lessen smoothing ability, the center-weighted median (CWM) [7] filter gives only positive integer weights to the central pixel. However, similar to SM filter the weighted median filters are performed across all pixels in an image: noise pixels and noise-free pixels. This significantly affects quality of the output image. The switching median filters is an common name for a group of filters that reduce number of pixels subjected to median filtration to those that are believed to be noise [8 ].Pixels identified as uncorrupted are left unchanged. The main part of each switching median filter is the impulse noise detection method. In this stage, different approaches have been incorporated to different switching median filters. For example, the pixel-wise MAD (PWMAD) [9] filter modifies MAD and uses it to subtract the impulse from noisy image as the noise detector; The adaptive central- weighted median (ACWM) [10] filter realizes noise detection by using the differences defined between the outputs of CWM filters and the current pixel of concern; The directional weighted median (DWM) [11] filter computes differences between the current pixel and its neighbors aligned with four main directions and chooses the smallest one as the reference to identify the noise pixels; The adaptive switching median (ASWM) 12] filter gives adaptive switching threshold, which is computed locally from image pixels intensity values in a sliding window, to 2013 Seventh International Conference on Image and Graphics 978-0-7695-5050-3/13 $26.00 © 2013 IEEE DOI 10.1109/ICIG.2013.14 40 2013 Seventh International Conference on Image and Graphics 978-0-7695-5050-3/13 $26.00 © 2013 IEEE DOI 10.1109/ICIG.2013.14 40 2013 Seventh International Conference on Image and Graphics 978-0-7695-5050-3/13 $26.00 © 2013 IEEE DOI 10.1109/ICIG.2013.14 40 2013 Seventh International Conference on Image and Graphics 978-0-7695-5050-3/13 $26.00 © 2013 IEEE DOI 10.1109/ICIG.2013.14 40 2013 Seventh International Conference on Image and Graphics 978-0-7695-5050-3/13 $26.00 © 2013 IEEE DOI 10.1109/ICIG.2013.14 40