• 作者： 王榮華; 林良達; 游銘德
• 作者服務機構： 國立臺灣海洋大學電機工程學系; 宏碁電腦公司
• 中文摘要： 本文發展一套可應用於還原脈衝雜訊影像之適應性神經模糊濾波器（histogram-based adaptive neuro-fuzzy filter，HANF）。利用輸入影像之像素分佈統計，我們提出可以估測出一組最佳（well-conditioning）初始隸屬函數（HIMF）之演算法。若與傳統隨機選定初始隸屬函數法比較，經吾人以實驗證明利用HIMF訓練HANF，可以獲致更快之收斂速度（因訓練時僅須微調）及較佳之影像還原效果（因HIMF在誤差曲面上較易避開區域性最小值）。其實，若估測之像素分佈統計與原始像素分佈誤差不大，則單以HIMF隸屬函數即足以獲得近似最佳影像還原效果。以PSNR（peak signal-to-noise ratio）作評比指標，HANF在脈衝雜訊率>20%情況下表現均比中值濾波器好。即當脈衝雜訊率≦20%時，HANF亦表現良好。 另外，本文亦針對像素分佈統計與最佳初始隸屬函數之關連性作理論探討及實驗，以瞭解HANF的概汎能力及適應性質。概汎能力方面是以幾張具相似像素分佈統計影像，隨機從中抽取一張訓練，再以此訓練後之HANF去還原本身暨其他影像，結果證明HANF有能力還原其他影像。甚且有些影像還原效果比接受訓練之影像優異，充份顯示HANF之適應特性。吾人認為由於HANF具良好的概汎能力以及適應性，足以減緩甚或免除“再訓練”之困擾，在未來實際應用（如：多媒體影音傳輸）上相當具有可行性。
• 英文摘要： This paper develops a novel adaptive filter, namely, the Histogram-based Adaptive Neuro-fuzzyFilter (HANF), which is a powerful tool for restoring images corrupted by impulsive noise. Effectivetraining of HANF is achieved by using an algorithm based on histogram statistics to estimate a set ofwell-conditioned initial membership functions (IMF). HANF is capable of fast learning because only fine-tuning is needed to complete the training process. In contrast to the inefficient random strategy, thehistogram-inferred IMF significantly reduces the likelihood of converging to an undesirable local minimum.In cases where training is prohibited, the histogram-inferred IMF alone can achieve near-optimal restoration.We show that HANF is superior to the traditional median filter (MF) in PSNR (peak signal-to-noise ratio)performance, especially when the impulsive noise rate > 20%. Furthermore, correlation between histogram statistics and HIMF is explored to study the adaptiveproperty and generalization capability of HANF. The generalization performance is tested by using imageshaving similar histogram statistics, from which only one image is selected at random to estimate IMFand to train HANF. The trained HANF is shown to be capable of recovering all the remaining images.With the histogram-inferred IMF, HANF essentially is free from the retraining problem. Comparisonsof various experimental results are given to show the superiority of HANF in restoration performanceand in practical implementation.
• 中文關鍵字： adaptive fuzzy filter; membership function; neural networks; image histogram; impulsive noise; generalization capability; retraining problem
• 英文關鍵字： --