针对区域范围内多幅待镶嵌影像之间的色彩差异问题，提出一种基于GPU的分块加权Wallis并行匀色算法。首先，根据变异系数对影像自适应分块并利用双线性插值确定每一个像素的变换参数，利用加权Wallis变换消除影像间的色彩差异。然后，为了控制区域整体的匀色质量，利用Voronoi图和Dijkstra算法确定影像间的处理顺序。最后，利用GPU技术进行并行任务设计并从配置划分、存储器访问和指令吞吐量等方面进行优化，提高算法运算效率。实验结果表明，本文方法既能有效地消除影像间色彩差异，又能消除影像间的对比度差异。与CPU串行算法相比，GPU并行算法显著减少了计算时间，加速比最高达到60倍以上。

Mosaicked remote sensing images that cover large areas are important in image analysis and application. However, different degrees of color and contrast differences are observed between images due to the influence of sensor and external factors, such as light and fog, which complicate image mosaicking. Therefore, eliminating the differences between adjacent images and ensuring consistent colors in the large area (i.e., color balancing) are becoming increasingly significant. The acquisition cycle of remote sensing data is shortened and the amount of data is increased dramatically with the development of the sensor technology. The changes bring challenges to the efficiency of color balancing of remote sensing images. The traditional serial processing model based on CPU also cannot meet the requirements of fast processing mass data to handle emergency response.
To solve the aforementioned problems, a parallel color balancing method based on adaptive block Wallis algorithm for image mosaicking was proposed. First, the images were adaptively divided into blocks depending on the coefficients of variation. Bilinear interpolation was used to determine the transformation parameters of each pixel, and the Wallis transform was adopted to eliminate the color differences between adjacent images. Second, Voronoi diagram was generated to determine the adjacent relation of images. Dijkstra algorithm was used to calculate the shortest path and determine the processing sequence for controlling the color consistency of the entire region. Finally, GPU technology was used to parallelize the proposed method for improving the efficiency. Bilinear interpolation and linear transformation are repetitive and dense computing tasks, which were directly assigned to each thread and executed simultaneously. The reduction method was adopted to parallelize the calculation of mean and standard deviation. Moreover, configuration, memory access, and instruction throughput were optimized to further improve the efficiency.
Two groups of experiments were implemented on orthoimages to verify the effectiveness and efficiency of the proposed method. Experimental results showed that the proposed method was superior to the traditional Wallis method and Inpho in visual effect and quantitative evaluation. Moreover, the highest speed-up of the proposed parallel algorithm based on GPU could be more than 60 times that of the serial color balancing method based on CPU.
The proposed method can effectively eliminate the color and contrast differences between adjacent images, thereby decreasing the difficulty in seamline detection. Meanwhile, the efficiency of the method is improved dramatically with the proposed parallel acceleration strategy. The performance of the proposed method is excellent in improving the quality and efficiency of color balancing and reducing the difficulty in image mosaicking. Moreover, the proposed method is sufficiently efficient to meet the requirements of fast color balancing of remote sensing images.