A scene text detection method based on long-term and short-term memory network

Abstract

The invention discloses a scene text detection method based on a long-term and short-term memory network. By using LSTM to model the feature, the invention obtains the spatial context relation of thefeature of the position and its upper, lower, left and right position, and then carries out text detection on the feature. Compared with the traditional method of using only local features, the scenetext detection method based on the long-term and short-term memory network provided by the invention can greatly reduce the error detection, can detect a lot of text information with inconspicuous local features, and reduces the omission of the text. The main innovation of the invention is to use LSTM to model the features horizontally and vertically, and to obtain the spatial context of the position and its upper, lower, left and right position features. Compared with the traditional scene text detection algorithm, this method can detect the text information with inconspicuous local featuresmore effectively, reduce the detection error rate and improve the detection accuracy.

Description

technical field [0001] The present invention proposes a scene text detection method based on a long short-term memory network (Long Short-Term Memory, LSTM). This method is a novel technique for scene text detection. Background technique [0002] Natural scenes contain a large amount of text information. In recent years, it has become a hot research direction for computers to accurately extract text information from natural scene images. Scene text detection is a key technology for many vision applications, such as it is widely used in unmanned driving, scene understanding, license plate recognition and other fields. However, due to the variability of fonts and scales in scene text, and the complexity of background and layout, it is difficult for traditional algorithms to obtain high performance. The method based on deep learning has become the mainstream method in this field, and its performance has also been excellent. However, the difficulty of this task still exists ob...

AI Technical Summary

Problems solved by technology

However, these connection criteria can only be applied to horizontal or near-horizontal text, and cannot detect text lines with a slightly larger slant angle; the second type is mainly based on convolutional neural network (Convolutional Neural Network, CNN) text detection algorithm, the algorithm The image is passed through CNN to obtain the convolutional feature spectrum, and then a candidate frame is generated at each position on the feature spectrum, and the position offset of the candidate frame is regressed to obtain the text area

[0004] These existing methods do not take into account the distribution characteristics of the scene text, that is, the scene text always appears in the form of text paragraphs or in the form of text lines,

Images