As the text detection in natural scenes is different from traditional object detection, using the region proposcal network (RPN) method proposed by FasterRcnn for text detection directly has some restrictions. On the one hand, because of variable length, background complexity, diversification of the text area and other factors, a greater receptive field design is required. On the other hand,
in the RPN training phase, there are a large number of false positives and missed detections in the selection of positive samples.
 
We propose a method based on perpendicular regional regression networks. Firstly, the Hough method is used to adjust the slope of the partial scene image. Secondly, in the training phase, based on the groundtruth box and the candidate box Anchor, the samples with an IOU value (intersection and union ratio) in vertical direction greater than a threshold, are selected as the positive sample. Thirdly, the positive samples in vertical direction are classified as regression. Finally, multiple adjacent Anchors are combined to form a text area. Experiments on the ICDAR2011 and ICDAR2013 data sets have a good detection result.