103學年度:基於GPU架構的高效能平行演化式影像雜訊過濾器設計

計畫名稱:基於GPU架構的高效能平行演化式影像雜訊過濾器設計
執行起迄:2014/08/01~2015/07/31
總核定金額:622,000元
中文摘要:雜訊過濾是將訊號中的雜訊減少或是移除,是影像或聲音處理過程中非常重要的環節。通常具適應性的過濾方法,其計算往往複雜度高,較無法滿足雜訊過濾實際時所要求的速度與效率。若將過濾方法實現在硬體上,可以利用硬體的優勢加速處裡的效率,提高其實用性;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,獲得具有適應性與硬體線路合理性的輸出。然而,設計演化式硬體影像濾波器的求解空間非常龐大(搜尋空間高達$10^{80}$以上),其計算過程非常耗時(單一解需數十小時),使得有限時間下所求得的過濾器設計效果受到限制,也是演化式硬體設計方式尚無法實用化的主要原因。本計畫依據演化式硬體影像濾波器設計所使用的CGP(Cartesian genetic programming)的特性,以圖形顯示處理器平台(GPU)有限的記憶體空間與執行緒數量的限制等特性,重新設計高效能的平行演化式運算演算法,將大量資料與運算工作切割成可以獨立並平行運算的子工作,並設計調整圖形顯示處理器處理執行緒的運算效能,將CGP的運算速度加以提升,來加速演化式硬體影像濾波器進行外部式演化的過程,以降低設計時間,並可以在相對時間內求得更多過濾器設計的解,提高其實用性。
英文摘要:Image denoising is to remove or reduce noise from the contaminated images and is an important research area in image processing. Evolvable Hardware (EHW) is a combination of evolutionary algorithm and reconfigurable hardware devices, that can change its architecture adaptively and produce flexible results for complex problems. One of the obstacles to realize an EHW-based method is its very long training time. The search space for solving an EHW problem may be up to $10^{80}$, which makes the realization of such EHW methods less possible. This project deals with the parallelism of EHW-based design of image filters using graphic processing units (GPUs). The design process is analyzed and decomposed into some smaller processes that can run in parallel. Pixel-based data for training and verifying EHW solutions are partitioned according to the architecture of GPU. Several strategies for deploying parallel processes will be developed and implemented. In this manner, the overall learning time can be reduced significantly for constructing EHW-based image filters.