在電腦視覺與影像辨識等領域之相關應用中，影像處理技術總是扮演著重要的角色，而影像前處理階段更會直接影響後續的結果與效能，本論文即是探討影像脈衝雜訊去除 (impulse noise removal) 與影像解析度增強 (image resolution enhancement) 等重要技術。較佳的前處理技術通常亦能提供更完善的結果給後續的處理步驟，然而，面對千變萬化的數位影像，卻找不到一個理論或是公式推導，可以直接適用於所有的情況，而眾人所熟知的模糊邏輯 (fuzzy logics) 相對於傳統數學模型，對於不確定性 (uncertainty) 更具有彈性，尤其是針對未知系統的近似與識別。因此，本篇論文提出了利用模糊邏輯的知識與規則，改善了傳統前處理技術的缺點，提供更好的結果。在雜訊去除中，本文所提出的模糊推論系統可得到一雜訊程度值，將可用於偵測影像中單一像素點是否為脈衝雜訊，若是則利用中值濾波器去除之，反之則保留。在影像解析度增強部份，本論文提出了兩種不同的空間域中影像內插演算法，適度地結合局部區域的梯度值與傳統線性內插的權重值，於影像中較為複雜、細節較多的區域，保留了更清楚的對比，而不至於破壞了影像中的邊緣與線條。第三章所提出的兩階段內插法可針對整數倍數的放大倍率使用，並採用邊緣的方向資訊，於內插後的影像可得較佳的方向性，延伸原始影像中的線條。第四章則是提供了更能廣為使用的內插法，可用於各種不同的放大倍數，不限制於整數倍數。由每一章中所提供的模擬結果可以觀察出本論文所提出之方法的確大大地改善了傳統方法的效果，不論是從每一次的實驗數據，亦或是實際去觀察套用於許多測試圖片的真實效果，本論文所開發的影像前處理技術確實在雜訊去除與影像的解析度增強，具有較好的效果，也直接影響了後續的影像處理技術。 Fundamental to numerous digital imaging applications and systems; image pre-processing techniques are widely researched in many engineering fields such as computer vision, industrial automation and pattern recognition. Some commonly used pre-processing techniques, including impulse noise reduction and image quality enhancement, are introduced in this dissertation. These pre-processing techniques that have good performance often provide highly-qualified results for the consequent imaging steps. However, neither mathematical definition nor derivation is given to deal with all types of images. It is very difficult to design and implement a critical image pre-processing method that is applicable to all possible image contents. As well-known that fuzzy logic is extensively adopted in locating and identifying models due to its capability to consider the uncertainties of unknown systems. Accordingly this dissertation tries to design several fuzzy-based image pre-processing methods such as impulse noise removal and image spatial resolution enhancement, so-called image zooming or enlargement, for improving the performance of those traditional methods. In impulse noise removal, the noisy pixels are first detected based on fuzzy logic and are then eliminated by a novel weighted median filter. In image zooming, we first propose a two-stage image interpolation approach with an integer-valued magnification factor. In addition, the parameters in employed fuzzy system are determined by using particle swarm optimization procedure. Finally, a more general image interpolation method and its modified version are presented to deal with the non-integer magnification factor. As demonstrated in the simulation results in their corresponding chapters, the proposed methods, including impulse noise removal and image interpolation algorithms, actually achieved good efficiency and accuracy.
