Text saliency-based scene text detection method

Abstract

The invention discloses a text saliency-based scene text detection method. The method comprises the following steps of initial text saliency detection, text saliency detailing and text saliency region classification. In the initial text saliency detection stage, a CNN model used for text saliency detection is designed, and the model can automatically learn features capable of representing intrinsic attributes of a text from an image and obtain a saliency map with consciousness for the text. In the text saliency detailing stage, a text saliency detailing CNN model is designed and used for performing further text saliency detection on a rough text saliency region. In the text saliency region classification stage, a text saliency region classification CNN model is used for filtering a non-text region and obtaining a final text detection result. By introducing saliency detection in a scene text detection process, a text region in a scene can be effectively detected, so that the performance of the scene text detection method is improved.

Description

technical field [0001] The invention relates to a scene text detection method. Background technique [0002] Scene text detection refers to locating the location of text regions in different scene images, such as road signs, store names and warning signs, etc., which is an important step in end-to-end scene text recognition. Efficient scene text detection results can help improve the performance of a large number of multimedia applications, such as mobile visual search, content-based image retrieval, and automatic sign translation, etc. In recent years, a series of international competitions on scene text detection have been successfully held, which has greatly promoted the research on scene text detection technology. However, due to many uncontrollable factors in the natural scene environment, such as different text sizes, color differences and complex backgrounds, scene text detection is still a very challenging problem in the field of computer vision. [0003] The first...

AI Technical Summary

Problems solved by technology

Traditional CTR extraction methods are usually based on sliding windows, stroke widthtransform (denoted as SWT) and maximally stable extremal region (denoted as MSER), so they do not make full use of the inherent intrinsic properties of the text itself. characteristics, resulting in the extraction of a large number of non-text candidate regions much more than the real text region, so that the task of subsequent non-text region filtering process is very difficult

Moreover, these traditional CTR extraction methods are sensitive to some external factors, such as illumination changes and image blur.

These will cause a part of the real text area to be extracted, resulting in a low recall rate

For example, the best recall rate of the currently published literature in the ICDAR2015 scene text detection library is only 0.83

Images