A Scene Text Detection Algorithm based on ResNet and Faster R-CNN

Text detection in natural scenes has become a research hotspot due to the continuous development of target detection technology. However, because of the influence of a scene text itself and external environmental factors, the conventional detection algorithm is not very effective. To solve the above problems, we propose in this paper a novel natural scene text detection algorithm based on the Faster R-CNN framework. We utilized a deep ResNet network instead of the VGG-16 network to extract, to a certain extent, more feature maps and improve the accuracy of the detection model. For candidate frames of different scales, we used different scales of pooling processing and classified the candidate boxes by cascading pooling features to achieve accurate positioning of the text. Finally, the Soft-NMS algorithm was used to post-process the region candidate frame. This method only attenuated the detection score of the non-maximum detection frame instead of deleting it, which effectively reduced the NMS threshold selection present in the conventional NMS algorithm. Consequently, false detection and missed detection were considerably reduced, thereby improving the accuracy of the text detection system. Moreover, experimental results revealed that our algorithm has significant advantages in F-value and detection accuracy over other widely used algorithms.