Tuberculosis is one of the most deadly diseases according to the World Health Organization. In 2008, 1.1-1.7 million people died and 8.9-9.9 million new cases were regis-tered. Currently, the most important tool of diagnosis is the direct examination of sputum smears. Since early diagnosis is the main strategy to control tuberculosis, faster methods of diagnosis are required. In this paper, an algorithm to detect bacilli of tuberculosis in microscopic images of Ziehl-Neelsen-stained sputum smears is described. First, a database of 1,340 images was created. The algorithm considered three stages: segmentation, feature extraction and classification. The seg-mentation stage was based on color empirical rules. The fea-ture extraction stage considered: Fourier descriptors, Hu moments and Zernike moments. The classification stage was based on a support vector machine. The algorithm reached 41.24% sensitivity. An improvement of this algorithm could represent a tool to rapidly identify risky sputum smears. © 2013 Springer.