結合小波轉換、方塊截短編碼、嵌入式零樹小波與資料探勘等技巧的新體積型影像壓縮方法

Abstract

本篇論文提出一個結合四種技巧的壓縮方法，本方法主要是結合小波轉換 (WT)、方塊截短編碼 (BTC)、嵌入式零樹小波(EZW)、資料探勘 (data mining)技巧而成。我們知道WT和BTC是兩種完全不同的壓縮方法，選擇結合這兩方法的原因是因為影像經小波轉換後，其所得的係數矩陣，再經由BTC原理加以分類後，將可獲得更高效應。但使用BTC分類後，相對的也會帶來位元圖產生之現象，然而將資料探勘技巧用於選擇位元圖上，可使所選的位元圖更是具代表性，且能有效的減少影像間所須傳送之位元圖數量。在前述一連串之動作後，所須傳送的係數明顯減少了許多，但如以固定長度位元來代表係數時，會使壓縮效果減低，所以在傳輸係數時，是使用了EZW演算法內的係數重建之方式來完成整個壓縮方法。而我們所完成的新方法經實驗証實確實有達到高壓縮比及高品質之優點。但在壓縮上所需的時間，比其它演算法來的差，由此可知若在壓縮時間上能有更進一步改善的話，這個新方法是個相當不錯的方法。本論文的主要貢獻有三，第一，利用BTC原理將小波轉換後之係數矩陣加以分類，使其所需之資料量減少。第二，利用資料探勘技巧有效的找出具代表性之位元圖及重覆之位元圖。第三，完成一個新的壓縮方法，實驗結果顯示此方法在壓縮上能夠達到不錯的成效。

This thesis proposes an image compression method which mainly combines four techniques including wavelet transformation(WT), block truncation coding(BTC), embedded zerotree wavelet(EZW), and data mining. They are two completely different types of compression method between WT and BTC. But, after the images were transformed through wavelet, the coefficient matrix being classified by BTC will result in higher effects; as a result, this thesis combines these two extremely different types of method, WT and BTC. However, using BTC to classify also brings about the problem of images generated by bitmap. Applying data mining to choose bitmap will not only make the chosen bitmap more representative but also effectively reduce the bitmap numbers during transforming the images. It is obvious that the above methods can decrease number of coefficient greatly. If we use the bits with fixed length to represent the coefficient, the performance of compression is poor. As a result, during transforming coefficients, the coefficients reconstruction method of EZW is adopted throughout the compression process. The new method in this thesis surely has high compression rate and quality. Nevertheless, the proposed method is more time-consuming if it compares to other algorithms. Therefore, if the problems of processing time can be improved, this new method is quite satisfactory. There are three major contributions of the present study. First, using BTC to classify the coefficient matrix obtained from wavelet transformation reduces data redundancy. Second, data mining techniques can effectively find out the representative and repetitive bitmap. Third, this thesis proposes a novel compression method and the experiment result shows the proposed method has good ability in compression.