Multi-language scene text detection and recognition method

Abstract

The invention discloses a multi-language scene text detection and recognition method. The method comprises: processing an input image through a text detector to obtain a series of text candidate boxes; adaptively generating a one-dimensional weight vector and a two-dimensional weight matrix through a feature selection and long-range dependency extraction module so as to pay more attention to channels from which text information is extracted and areas containing texts, and acquiring global information by capturing long-range dependency; and by the operation of the feature enhancement module, enabling the network to have better distinguishing performance on the text / non-text, thereby reducing false alarms; finally, predicting a series of text candidate boxes by using a plurality of convolution respectively; and carrying out text recognition and text category recognition on the text candidate boxes subjected to threshold processing and zooming through a text recognizer and a text category recognizer to obtain text contents and text categories. The method has high text detection and recognition performance and is suitable for multi-language application scenarios.

Description

technical field [0001] The invention relates to the technical field of text detection and recognition, in particular to a multilingual scene text detection and recognition method. Background technique [0002] Scene text reading refers to the detection and recognition of all text contained in natural scene images. It has many applications in image retrieval, scene understanding, automatic driving and text translation. At present, many research methods only do text detection or text recognition alone. There are some methods that can perform text detection and recognition at the same time, but they are mainly for text in one language (for example, English or Chinese), that is, these methods cannot be applied. in multilingual scenarios. [0003] Furthermore, these methods only use local operations such as convolutional neural networks and recurrent neural networks, which do not capture long-range dependencies. Furthermore, these methods generally use online hard case mining a...

AI Technical Summary

Problems solved by technology

At present, many research methods only do text detection or text recognition alone. There are some methods that can perform text detection and recognition at the same time, but they are mainly for text in one language (for example, English or Chinese), that is, these methods cannot be applied. multilingual scene

[0003] Furthermore, these methods only use local operations such as convolutional neural networks and recurrent neural networks, which do not capture long-range dependencies

Furthermore, these methods generally use online hard case mining algorithms to reduce the false positive rate of the network, but their efficiency is low.

Finally, existing methods only use Connectionist Temporal Classification (CTC) or attention-based decoders to decode input sequences into text, making text recognition performance low

Images