High-Throughput Median Filtering for Large Kernel Sizes on CUDA

In this paper, we propose a novel approach to median filtering, a widely employed technique for mitigating specific types of noise in images and signals. While median filtering is effective, its computational demands, especially with larger filter sizes, pose challenges for real-time applications. To address this, we present a highly efficient, optimized, and parallelized median filtering algorithm tailored for CUDA platform. Leveraging a histogram-based method and operating on 8-bit data, our approach outperforms sorting-based alternatives, particularly for larger filter sizes. Quantitative experiments demonstrate substantial performance improvements, with our algorithm achieving speedup up to 20x and 59x compared to the second best GPU-based and CPU-based algorithms for mid and large filter sizes, respectively. This significant enhancement in processing speed makes our algorithm a compelling choice for real-time applications where rapid noise removal is paramount, thereby extending the practical utility of median filtering in various domains.