現今的資訊化時代中，影像存取皆使用數位影像紀錄，而數位影像容易受到雜訊干擾，以致影像畫質下降，無法辨識原始影像中之資訊；因此，如何有效移除雜訊並修復數位影像，是一項十分重要之研究領域。本文將基於方向型權重中值濾波法，使用深度學習之技術，優化方向型權重中值濾波法之效能。我們採用乾淨影像像素的變動方向作為訓練目標集，再將同一張影像加入椒鹽雜訊，受雜訊干擾影像之像素間的變動方向作為訓練集，訓練深度學習類神經網路，以此訓練好之網路用估測受雜訊干擾像素的變動方向，並且利用此方向的乾淨像素於修復雜訊干擾像素。在處理高雜訊密度干擾的影像時，可參考之乾淨像素不足，本文採用疊代方式做重複處理，直至重建清晰之原始影像。經實驗結果證實：本研究之方法確實可在任何雜訊密度下，有效移除數位影像中之椒鹽雜訊，重建高品質之影像。
An image would be interfered with salt and pepper noise, causing the quality of the image to be degraded. An effective method to remove noise and restore the image is important. Although a Directional-Weighted-Median (DWM) filter can effectively restore the edge information in image denoising, this filter only considers the local properties in defining the weighting factors for the neighboring pixels. The global information of the image is ignored. In this thesis, we employ a deep-learning neural network (DNN) to define the direction of pixel variation for a noisy pixel. Hence, the noisy pixel is restored by the noise-free neighboring pixels on the direction defined by the DNN. Because the DNN is trained by using a noisy image and the corresponding noise-free image as the target image, the global properties of the noisy image and its noise-free image are considered. Therefore, the proposed DNN-based DWM filter can improve the performance of the DWM filter significantly. In the cases of heavy noise corruption, an iterative processing on noisy pixels is performed. The experimental results reveal that the proposed filter can effectively remove salt and pepper (SAP) noise in an image corrupted by various noise densities.
