Noise Reduction with Non-Local Mean and Hierarchical Edge Analysis

消除雜訊是數位影像處理中極為重要的一個步驟，尤其在手機及視訊相機中的雜訊特別明顯。而雜訊的出現會使得影像的品質大幅度的降低，而造成使用者很難拍出一張能令其滿意的相片。且有時也會由於其他數位影像處理而使得雜訊增加，如數位防手震-其用調高ISO值來達到防手震的目的，但也會使得雜訊大增，因此就需要使用去雜訊的方法。此篇論文中，我們使用非區域平均化及階層式邊分析的方法來做到將雜訊消除。此方法既保留了非區域平均法在邊區域保留細節的優點且在非邊部份有消除雜訊的效果Noise reduction is a very important step in digital image processing. The noise problem is serious in cell phone and video camera. The noises result in lower image quality. Moreover, users can not take a satisfactory photo. Furthermore, other digital image processing may also increase the noises. For example, digital image stabilization by raising the ISO speed will increase the noise level.n this thesis, we use non-local means and hierarchical edge analysis to reduce noises. Our method not only preserves the detail of the edge area with non-local means but also retains noise reduction on non-edge area with hierarchical edge analysis.口試委員會審定書 i 謝 ii 要 iiibstract ivhapter 1 Introduction 1.1 Introduction to Noise Type 2.2 Noise Models 4.3 Noise Level Measurement 4hapter 2 Noise Reduction Algorithm 6.1 Gaussian Bilateral Filter 6.2 Y. L. Huang’s Bilateral Filter 7.3 Wavelet Transform 9hapter 3 Non-Local Means and Hierarchical Edge Analysis Methods 10.1 Non-Local Means Method 11.1.1 Non-Local Means Equation 12.1.2 Gaussian Kernel 13.2 Hierarchical Edge Analysis 14.2.1 Construct Hierarchical Frequency 14.2.2 Determining Edge Pixels (104) 16.2.3 Determine Edge Pixel Direction (112) 19.2.3 Smooth the Edge Pixels (114) (116) 20.2.4 Combine the Different Frequency Images (118) (120) 21hapter 4 Our Proposed Method 23.1 Drawbacks 23.1.1 Non-Local mean’s drawback 23.1.2 Drawback of Hierarchical Edge Analysis 25.2 Our Proposed Method 26.2.1 Color Space 28.2.2 Hierarchical Edge Finding and Non-Edge Noise Removal 28.2.3 Edge Noise Reduction 30hapter 5 Experimental Results, Conclusion, and Future Work 31.1 Experimental Results 31.1.1 Experiment Environment 31.1.2 Gray Level Image Experiment 31.1.3 Color Image Experiment 35.2 Conclusion 48.3 Future Work 48eference 4