- 學位論文
基於GPU之時域DPCM應用於即時倒車彩色影像壓縮編碼法
Real-time Time Color Video Compression Based on GPU Time-domain DPCM
摘要
近年來，行車紀錄器與倒車影像系統的普及，使人們在行車安全上更具有保障，然而因生活品質的提升，人們對於影像品質的要求也越來越高，導致影像壓縮編碼演算法的複雜度也隨之提高，造成使用者必須在「影像即時性」和「影像品質」間進行抉擇，無法兩者兼顧；因此，本研究將透過高效能平行運算技術(Graphics Processing Unit，GPU)來提升系統的處理速度，讓倒車系統能在最小壓縮失真的情況下即時傳送影像給使用者。 本論文採用的倒車影像編碼為「誤差脈衝編碼調變」(Differential PulseCodeModulation，DPCM)，將影像分為灰階Y與色彩UV，並分別利用不同的縮放技術來進行處理，其中因為Y對於影像輪廓的呈現較為重要，所以其縮放比例較小，反之，UV的縮放比例則較大，接著對UV進行位元平面縮減，最後再將所有的影像運算透過GPU來進行加速。 實驗結果證明平行化「時域DPCM影像視訊壓縮編碼」可大幅降低運算時間，並在維持影像品質下提高每秒顯示幀數(Frames Per Second，FPS)，最後再搭配統一計算架構(Compute Unified Device Architecture，CUDA)所提供的串串流(Stream)技術執行非同步資料傳輸 (Asynchronous Data Transfer，ADT)以減少整體傳輸的時間。經實驗顯示影像大小為1440*1088像素值(pixels)時，GPU對比中央處理器(CPU)色彩空間Y解碼的執行時間的加速可高達44.8倍。
並列摘要
In recent years, car recording and reversing video system has become universal, so that people are safer on road. However, due to the quality of technology improvement, image quality requirements are also increasing, there for complexity of image compression algorithm has increased.Causingto choose between "instant image" and "image quality"；Therefore, this study will base on high-performance parallel computing technology,to improve the processing speed of the system, so reversing system can immediately send images with minimal compression. In my thesis the video encoding is based on “Differential Pulse Code Modulation, DPCM”, Y image into grayscale and color UV, respectively, using different scaling techniques for processing, in which Y is more important so that outline for the presentation of images scale is smaller, another word the UV scale is greater, then carry on UV-bit flat down, then all of the final image through GPU computing to accelerate. Experimental results show that the parallel of " DPCM video compression image coding" can significantly reduce the computation time and maintain image quality under improved display FPS(frames per second), andwithimplementation of Stream technology that provide by CUDA (Asynchronous Data Transfer, ADT) to reduce the overall transmission time. The experiment is displayed by the size of 1440 * 1088 pixels, GPU accelerated CPU execution time comparison can be up to 44.8 times faster.