Text detection in document images plays an important role in optical character recognition systems and is a challenging task. The proposed text detection method uses self-adjusting bottom-up segmentation algorithm to segment a document image into a set of connected components. The segmented connected components are then described in terms of 27 features and a machine learning algorithm is used to classify these components as text or non-text. We have collected a dataset (called ASTRoID), which contains 500 images of text blocks and 500 images of non-text blocks in order to test the method. We empirically compare performance of the proposed text detection method with seven different machine learning algorithms; the best performance is obtained with the radial support vector machine.