Efficient Scalable Median Filtering Using Histogram-Based Operations.

Area of Science:

• Computer Vision
• Image Processing
• Parallel Computing

Background:

• Median filtering is crucial for image noise removal but existing parallel implementations are often inefficient due to sorting.
• Limited efficient parallel median filtering solutions exist for multi-core systems and accelerators.

Purpose of the Study:

• To develop a novel, non-sorting-based parallel median filtering algorithm.
• To improve computational efficiency and reduce image access for median filtering.
• To implement and evaluate the algorithm on both CPU and GPU (CUDA).

Main Methods:

• Introduced a new parallel median filtering algorithm utilizing efficient histogram-based operations.
• Implemented the algorithm for Central Processing Unit (CPU) and Graphics Processing Unit (GPU) using NVIDIA's CUDA.
• Compared performance against leading CPU and GPU median filtering implementations.

Main Results:

• The CPU implementation demonstrated near-perfect linear scaling on a quad-core system.
• The GPU implementation achieved significant speedups, orders of magnitude faster than existing GPU methods for mid-size filters.
• Comparison-based methods remain preferable for very small filter kernels.

Conclusions:

• The new histogram-based median filtering algorithm offers superior performance, especially for larger filters on parallel architectures.
• The algorithm provides an efficient, open-source solution available in the OpenCV library.
• This non-sorting approach represents a significant advancement in parallel image processing techniques.