Locating Texts in Images by Fully Convolutional Network with Bounding Text Labeling

Texts in images are often viewed as the regions of interest. Locating such areas for further analysis may help to extract image-related information and facilitate many interesting applications. Considering that pixel-based image segmentation is effective in identifying the areas containing texts, we propose a text detection scheme based on the Fully Convolutional Network (FCN), which employs Feature Pyramid Network (FPN) and Atrous Spatial Pyramid Pooling (ASPP) to effectively mark the pixels related to texts in an image. A comparatively efficient labeling approach is adopted to group related letters/characters and separate different words by oval-shaped text labeling masks with boundaries. The experimental results demonstrate feasibility of the proposed text-detection scheme when compared with the state-of-the-art text detection approaches.