부산대학교 도서관

Connected Components Labeling (CCL) is a fundamental task in binary image processing. Since its introduction in the sixties, several algorithmic strategies have been proposed to optimize its execution time. Most CCL algorithms in literature, including the current state-of-the-art, are designed to work on an input stored with 1-byte per pixel, even if the most memory-efficient format for a binary input only uses 1-bit per pixel. This paper deals with connected components labeling on 1-bit per pixel images, also known as 1bpp or bitonal images. An existing run-based CCL strategy is adapted to this input format, and optimized with Find First Set hardware operations and a smart management of provisional labels, giving birth to an efficient solution called Bit-Run Two Scan (BRTS). Then, BRTS is further optimized by merging pairs of consecutive lines through bitwise OR, and finding runs on this reduced data. This modification is the basis for another new algorithm on bitonal images, Bit-Merge-Run Scan (BMRS). When evaluated on a public benchmark, the two proposals outperform all the fastest competitors in literature, and therefore represent the new state-of-the-art for connected components labeling on bitonal images.