基于图形处理器的相位校正稳像

摘要

Because motion estimation algorithm occupies 60-80% of the entire computation in image stabilization algorithm, and the difficulty increases with the improvement of image resolution and real-time stability of continuous image sequences, a GPU-based phase correction strategy for stabilization phase correction algorithm is presented. Based on mathematical models, a unified parallel architecture of image processor and pixel single-threaded mode are used, the phase of motion estimation overlapping composite structure is designed, improving the performance of the parallel thread synchronization. The computation time of parallel FFT is reduced. At the same time, change the scale of parallel threads, adjust the graphics processor storage, improve memory access performance data and motion estimation phase of concurrency, and estimate a stable rotation and translation etc. Required for video in order to achieve fast and stable gray scale resolution of 1 k × 1k purpose of continuous video. Experimental results show that the average time is about one frame stable 10ms, which effectively enhance the efficiency of video stabilization, and the possibility of image stabilization phase engineering applications.%由于稳像算法中的运动估计占据整个算法60％～80％的运算量,并随着图像分辨率的提高,实时稳定连续图像序列的难度不断增加.本文提出一种基于图形处理器的相位校正策略,根据相位校正稳像算法教学模型,利用图像处理器统一并行架构与像素单线程模式,设计交叠复合相位运动估计结构,提升并行线程同步性能,缩短FFT并行运算时间；同时通过改变并行线程的尺度,调整图形处理器的存储方式,提升数据存储器的访问性能,改进了相位运动估计的并发性,可高品质估算视频需要稳定的旋转,平移等变量,以达到高速稳定1k×1 k分辨率的灰度连续视频的目的.实验证明,平均稳定一帧图像时间约为10 ms,有效提升了视频稳定的效率,增强了相位图像稳定工程应用的可能性.