計畫名稱：模糊類神經建模於高效能演化式硬體影像雜訊過濾技術之研究
執行起迄：2013/08/01~2014/07/31
總核定金額：650,000元
中文摘要：雜訊過濾是將訊號中的雜訊減少或是移除，是影像或聲音等訊號處理過程中非常重要的過程。使用固定模式的雜訊過濾器進行雜訊去除，因為方法固定，通常效果有限。具有適應性的過濾方法，其計算過程往往複雜度較高，也較無法滿足雜訊過濾實際時所要求的速度與效率。將軟體形式的適應性過濾計算方法實現在硬體上可以以加速過濾的實用效率，但硬體的複雜度與製作成本也隨著計算方法的複雜而大幅增加。演化式硬體結合演化式演算法及可重構式硬體，可以依照外在環境的變化，以演化的方式調整內部電路結構，獲得具有適應性與靈活性的輸出。要設計演化式硬體的過濾機制，有兩個重要的議題須考慮：一、有效的適應性機制；二、過濾成效與製作成本的平衡。通常這兩個議題是互相衝突的，成熟的適應機制往往需要引用複雜的計算程序，雖然可以提升過濾成效，但通常也會增加實作成硬體時的線路複雜度。本計畫將致力於適應性演化式硬體影像雜訊濾波器的設計問題，以提升硬體影像雜訊濾波器的成效為目標，提出兩年期計畫：分別為第一年：「Type-2模糊類神經建模技術於影像雜訊類型自動建模之研究」依據根據像素之間的相似度與差異性，以Type-2模糊類神經技術進行有效的雜訊系統建模，作為演化式硬體設計時的有效架構；第二年：「高效率的演化式多目標最佳化線路合成演算法的設計」發展同時考慮過濾正確性與線路複雜度的高效率硬體線路合成演算法。目前計畫通過為一年期計畫，故執行與「Type-2模糊類神經建模技術於影像雜訊類型自動建模之研究」的部分。
英文摘要：Denoising is to remove or reduce noises from the contaminated images and is an important research area in image processing. Denoising methods can work well if the noise models can be known in prior. However, it may lack of flexibility and adaptability when un-modeled types of noises are encountered. Evolvable Hardware (EHW) is a combination of evolutionary algorithm and reconfigurable hardware devices. EHW can change its architecture adaptively and produce flexible results for complex problems. We propose a two-year project regarding the design of efficient and effective EHW-based image filters using the neural fuzzy modeling techniques and multi-objective optimization methods. In the first year, images for training EHW are analyzed and noise models are constructed automatically by type-2 neural fuzzy modeling techniques. Each EHW filter can be trained efficiently and effectively by only homogeneous data that are categorized and reduced in size by the neural fuzzy models. In the second year, efficient multi-objective optimization methods will be developed that consider both the filtering accuracy and complexity of hardware circuits. In this manner, the overall learning time can be reduced with less circuit complexity for constructing EHW-based image filters. This project was granted by NSC as a one-year project. This report presents the results of the first-part of the project.