EAMF: Efficient Approximate Median Filter Hardware Architecture Using Nonsorting Method for Salt-and-Pepper Noise Removal

This paper proposes four architectures of approximated median filters, designed to effectively remove salt-and-pepper noise in image introduced during acquisition, transmission or storage. The process begins with noise pixel detection, followed by their replacement with an approximated median value. The 8-bit input pixel data are divided into multiple slicing parts, ensuring that the last part of slicing always consists of either 3 or 4 bits. The approximated median is computed by determining the exact partial median for each slicing part except the last slicing part, where an approximation is applied. Among the four architectures, two architectures operate on nine input pixels, while the other two uses eight input pixel data. The image quality assessment of the proposed architectures is performed in terms of PSNR and SSIM using various grayscale image datasets. Notably, one of the proposed architectures achieves nearly identical image quality while occupying an area of 5192[Formula: see text] [Formula: see text] and consuming 0.375[Formula: see text]mW of power, leading to over 50% savings compared to state-of-the-art methods.