Hardware-friendly architecture for a pseudo 2D weighted median filter based on sparse-window approach

Stereo matching, which is conventionally used for three-dimensional (3D) information acquisition through cameras, is the most actively studied subject in computer vision. To obtain sophisticated 3D information, refining the disparity map in stereo vision is important. The weighted median filter (WMF) is extensively used to eliminate outliers in post-processing. To this end, various studies have implemented WMFs in hardware for real-time processing. Among them, the separable weighted median filter (sWMF) vertically and horizontally separates a two-dimensional WMF into two one-dimensional WMFs to reduce hardware resource usage. Herein, we propose a hardware architecture that can reduce the hardware resource usage of the sWMF by applying the sparse-window approach, which is a method of creating a window by selecting pixels sparsely. This approach makes it possible to reduce drastically the number of elements to be computed. Although the proposed architecture has an insignificant disparity error rate, similar to that of the sWMF, it saves 33% slice lookup tables (LUTs) and 69% slice registers when using a window size of 37 × 37 pixels as the synthesis result on the Xilinx XC7K325T FPGA. When a window size of 49 × 13 pixels with the best performance is used, the proposed architecture uses 7335 slice LUTs, 4126 slice registers, and 21 block RAMs. The proposed architecture operates at frequencies of up to 167.95 MHz; hence, it can operate in real time. The proposed WMF architecture is suitable for application in embedded systems and low-resource environments as it is hardware-friendly.