Impulse noise removal on an embedded, low memory SIMD processor

Vector median filters efficiently reduce noise while preserving image details. However, their high computational complexity for color images makes them impractical for real-time systems. We propose new computationally efficient filtering algorithms, called index mapping filters (IMF). These filtering algorithms are accelerated by implementing them on a massively data parallel processor array. In addition to greater computational efficiency, these algorithms result in robust noise reduction of corrupted color images. Analyses of mean square error, signal-to-noise-ratio, and visual comparison metrics indicate that IMF are competitive with the vector median filter (VMF) in their ability to correct impulse noise in color images. These algorithms are implemented on a SIMD processor array being developed for high efficiency, high-performance portable products. Executing on a 4096 node SIMD chip operating at 50 MHz, IMF 3×3 window applied to a 256×256 color image would take 442 microseconds (22104 clock cycles) for index mapping distance filter (IMDF) and 408 microseconds (20415 clock cycles) for index mapping median filter (IMMF).