A quadrilateral scene text detector with two-stage network architecture

Highlights：

• We propose a novel quadrilateral regression algorithm for generating quadrilateral proposals and text detections.

• We introduce a dual-branch structure of a quadrilateral detection head and an additional rotated rectangle detection head to train the detection head at the second stage.

• We design a novel weighted RoI pooling module with learned weight masks to pool the feature.

• We propose an accelerated NMS algorithm for quadrilateral text proposals and detections.

• We integrate the above four novel modules into the two-stage architecture of Faster R-CNN, and achieve state-of-the-art results on multiple public datasets.