Language identification method of scene text image in combination with global and local information

Abstract

The invention discloses a language identification method of a scene text image in combination with global and local information. Basic features of a character image are extracted, and then global andlocal feature representations are extracted respectively; the global extraction branch uses global maximum pooling to express the whole graph as a vector, and category score prediction is carried out;probability prediction is performed on the local blocks of the image by the local aggregation branches respectively, and then the series of probability distributions are combined to obtain a categoryprediction score of a local level; and finally, global and local prediction scores are dynamically fused according to the branch prediction conditions to obtain a final identification result. According to the method, overall features and local differentiated features of the character images are noticed at the same time, and end-to-end training can be achieved in one step. Compared with an existing technology utilizing local features, the method has the advantages that the local differentiated features can be accurately extracted, excellent effects are achieved in the aspects of accuracy, operation efficiency and universality, and high practical application value is achieved.

Description

technical field [0001] The invention belongs to the technical field of computer vision, and more specifically relates to a language recognition method of a scene text image combining global and local information. Background technique [0002] The task of image language recognition is to identify the language of the text contained in a given text image. With the development of globalization, the task of language recognition plays an increasingly important role in today's multilingual systems. It can be seen from the results of the ICDAR MLT competition that the same model has great differences in the detection and recognition of characters in different languages. In many cases, we need to choose which language model to use for further text detection or recognition, and language recognition can help us provide automated choices. In the end-to-end multilingual text detection and recognition system that may appear in the future, the text language can be used as an important in...

AI Technical Summary

Problems solved by technology

First, they considered the contribution of all high-dimensional features to the recognition results, but if there are many non-discriminative characters, the resulting redundant unimportant features will make the truly important features less significant, or Discriminative features are underutilized due to numerical disadvantages

Secondly, the method of directly cutting out a series of image blocks from the original image will cause some image blocks to be unrecognizable at all, which will interfere with the network's prediction of the entire image, and will generate redundant and unnecessary calculations.

Images