A practical parallel implementation for TDLMS image filter on multi-core processor

Abstract

In this study, parallel implementation of adaptive image filtering algorithm based on two-dimensional least mean square method (TDLMS) where the weights are continuously adjusted during filtering was realized by proposed design considerations. Despite its strictly sequential structure, the effect of a pixel on weights vanishes as the filter mask progresses. Based on this property, the load of filtering algorithm is allocated to threads by splitting the input image into sub-blocks. Due to the discontinuities, the crossing distortions between sub-blocks were eliminated using weight synchronization with the neighbor sub-block. Performance evaluations for various sizes of images were realized on a computer with multi-core processor using open multiprocessing library. In spite of the sequential nature of the algorithm, results show that the parallel implementation provides significant improvements in terms of both speedup and parallel efficiency.