Authors: Gem AI

Description: Text detection algorithms focus on detecting the existence and location of text. Specifically, the task is to output the outline of each text target but is not concerned with identifying the specific text content. Mainstream text detection algorithms can be divided into two categories. The first type is very similar to target detection, and the second is based on the idea of ​​image segmentation.
Among these methods, regression plays an important role in Bbox acquisition, but it is not essential, since the text/non-text prediction itself can also be viewed as a segmentation. semantics containing the complete location information. However, instances of text in scene images are often very close together, making it difficult to separate them through semantic segmentation. Therefore, instance segmentation is needed to solve this problem. This paper proposes a PixelLink scene text detection algorithm based on version segmentation. The first version of the text is segmented by linking pixels in the same version together. Then extract the bounding box text directly from the segmentation result without performing position regression.