一個改良式切換中值濾波器由高污染影像移除脈衝雜訊

An Improved Switching Median Filter for Impulse Noise Removal from Extremely Corrupted Images

本論文對於切換式中值濾波器提出一個新的脈衝雜訊偵測演算法，此演算法結合了兩種脈衝雜訊偵測器，第一種脈衝雜訊偵測器是產生六個方向的迴旋積遮罩，其利用一維拉普拉斯運算元計算後找出最小值，來判斷目前考慮的像素是否有受到脈衝雜訊的污染以保留邊界資訊，第二種脈衝雜訊偵測器是基於邊界差異的雜訊偵測器 （boundary discriminative noise detection） 簡稱BDND，BDND 演算法主要是將目前考慮的像素及其鄰近的像素利用兩個計算出來的邊界值，將這些像素群分成三群，分別是低濃度像素群、中濃度像素群、高濃度像素群，如果目前考慮的像素被分類到低濃度像素群或是高濃度像素群，則判定目前考慮的像素受到脈衝雜訊的污染，反之若被分類到中濃度像素群，則判定目前考慮的像素沒有受到脈衝雜訊的污染。脈衝雜訊偵測完成後將產生兩個二位元的脈衝雜訊圖，如果目前考慮的像素皆被判定受到脈衝雜訊的污染，我們使用中值濾波器將脈衝雜訊移除，反之則以原像素輸出結果。實驗結果顯示，本論文所提的切換式中值濾波演算法，能有效的移除脈衝雜訊，並保留絕大部份的細節，另一方面在0%到70%這高範圍的脈衝雜訊濃度中，所提切換式中值濾波演算法相較於其他切換式中值濾波演算法，不論在細節、線條、或是邊界部份等的脈衝雜訊移除，都有相當優秀的表現。

A new technique of impulse noise detection for switching medianfilters is presented in this paper. It consists of two impulse noise detectors.The first detector is the impulse noise detection method called theboundary discriminative noise detection (BDND), its target is to remove noises from the extremely corrupted images. The second detector is basedon the minimum absolute value of six convolution kernels obtained usingone-dimensional Laplacian operators. It can preserve detail whiledenoiseing.The drawback of the current convolution-based median filters is that ifthe pixels at the two end points of each edge are determined as noises, theythen are removed and smoothed out. Two new convolution kernels areproposed to solve this problem in this paper. However, convolutionkernelbased filters usually obtain bad results when the images arecorrupted by high density impulse noises. For overcoming this problem,the BDND algorithm is adopted. It first classifies the pixels of a localizedwindow, centering on the current pixel, into three groups—lower intensityimpulse noise, uncorrupted pixels, and higher intensity impulse noise.The center pixel will then be considered as “uncorrupted,” provided that itbelongs to the “uncorrupted” pixel group, or as “corrupted”. The twoboundaries that discriminate these three groups are determined so as toyield a very high noise detection accuracy.Two binary impulse noise maps are constructed from the above twodetectors. If both of them classify the current pixel as noise, a median filteris applied to remove the impulse noise. Otherwise, the pixel is determinedto yield image details, it is preserved. Experimental resultsshow that the proposed switching median filter can effectively restoredigital images extremely corrupted by impulse noise while preservingdetail. A comparison of the proposed method and other approaches in theliterature clearly shows that the proposed method substantially outperformsother schemes.