Text detection in medical images using local feature extraction and supervised learning

In this paper, a novel method to automatically detect the texts embedded in medical images is proposed. Specific local features for texts in medical images, such as local edge density, local intensity contrast, and connectivity, are defined and extracted to find out the candidate text regions. Then the histograms of oriented gradient (HOG) for all candidate regions are calculated. With both the HOG features and the aforementioned local features, an adaptive boosting (AdaBoost) classifier is used to discriminate the texts from non-text structures. Experimental results show that the proposed method has better text detection performance compared with previous methods. It can preserve the text information and eliminate the obstruction caused by different sources. The detected texts can provide additional information in many applications such as medical image retrieval.