Method and apparatus for text detection and positioning in natural scene based on focus loss function

Abstract

The invention discloses a text detection and positioning method and a device under a natural scene based on a focus loss function. Firstly, the annotated data is preprocessed, then a text detection and location network is constructed, and then the focus loss function is used as a part of the training process loss function, and then the natural scene images to be detected are detected. The method adjusts the existing tagging to make the tagging more suitable for the designed character detection network. FCN-based network combine multiple convolution layers to make them more consistent with thattask of text detection; by introducing the focus loss function to balance the positive and negative samples in the training process, the detection accuracy is improved. The invention can obtain the effect of high precision and high recall on character detection and positioning.

Description

technical field [0001] The invention belongs to the technical field of computer vision, and in particular relates to a method and a device capable of accurately locating text regions in natural scene pictures. Background technique [0002] There are many ways for human beings to disseminate information. As the carrier of information dissemination, text itself directly contains rich semantic information. In natural scenes, text is ubiquitous. Whether it is shop signs, traffic signs, or even street advertisements, posters, etc., words are used to convey information. Accurately locating and recognizing text areas from natural scenes can help machines better understand the semantic content of scenes, and is helpful in many fields. For example, in the field of street view recognition, recognizing the text on building plaques will help us better understand street view information; in the field of assisted driving, recognizing the text on traffic signs will help us better assist ...

AI Technical Summary

Problems solved by technology

The third is that the picture quality of the processing objects is different

However, there is a big difference between the text in the natural scene and the object in the object detection, mainly reflected in the fact that the text area may be longer, and the direct use of the object detection method is not targeted and the effect is not good

Therefore, it is still a big challenge to design a reasonable and efficient text detection method according to the characteristics of text.

Images