HTD: A Fast Human-centered Text-locating Method for Auxiliary Reading

As computer vision and human-computer interaction technologies are getting mature, vision-based auxiliary text reading has become the mainstream method to optimize learning and reading experience. To overcome the drawbacks of the common multi-step detection way and further improve the text reading experience, we propose a new human-centered fast text-locating method, named HTD. It transforms the relatively time-consuming serial procedure to a hybrid object detection problem, utilizing the hand-text hybrid object detection model to instantly locate the text of interest to readers. Considering the particularity of this task, we choose to perform the weighted NMS to infer final predicted bounding box. To verify the effectiveness of the proposed method, we tested the performance on the homemade hand-text hybrid object detection dataset HTDD. The experiment indicated that our proposed method is more efficiency and can produce satisfying results, which could meet the needs of human-computer interactive auxiliary reading in text reading scenarios.