Design of a Text Detection System via Hypothesis Generation and Verification

Text-detection is a well-established topic within the document recognition domain, but detecting text in natural scenes has proved to be considerably harder due to the large variability of text appearances, and due to dominating clutter. In this paper, we begin with a practical model for text appearance and use it to motivate the design of a text detection system consisting of two stages: a hypothesis generation stage and a hypothesis verification stage. For hypothesis generation, we adapt the Maximally Stable Extremal Regions (MSER) algorithm for the task of text detection by a combination of judicious parameter selection and a computationally efficient multi-scale analysis of MSER regions. For hypothesis verification, we design simple and fast image operations that produce features highly discriminative of text, which require minimal training, whose performance bounds can be specified, and which can be tuned in terms of intuitively understandable properties of text such as font, stroke-width, spacing, etc. Our final contribution is a new dataset of images of urban scenes annotated for text. The annotations include not only bounding boxes but also a richer set of text attributes. There are about 6000 text items in the dataset making it to our knowledge, the largest, most challenging, and most detailed text dataset for urban scenes. We show improved performance than state-of-the-art on the well-known ICDAR datasets and results on our new dataset.