Semantic segmentation-based text detection method for scene of any shape

Abstract

The invention discloses a semantic segmentation-based text detection method for a scene of any shape, wherein the method comprises the following steps of S1, constructing a semantic segmentation-basedtext detection network model for a scene of any shape; S2, according to the overall target loss function, minimizing the overall loss function by using a back propagation algorithm and stochastic gradient descent optimization, and carrying out iterative training on the model designed in S1; S3, using a step-by-step scale expansion method, performing scene text detection and recognition accordingto the model trained in the S2. For the text detection problem of text instances which are in any shape and are close to each other in a natural scene, the invention provides a semantic segmentation-based method for creatively utilizing a multi-kernel step-by-step scale expansion method to detect the text, so that the positions of the text blocks are detected more accurately.

Description

technical field [0001] The invention relates to the field of computer vision, in particular to a method for detecting text of arbitrary shapes in a scene based on semantic segmentation. Background technique [0002] With the development of convolutional neural networks, scene text detection has achieved rapid development, and it is currently widely used in geolocation, real-time translation, and assistance for the blind. But the detection of scene text is different from traditional optical character recognition (OCR), which is more challenging due to multi-directional, curved and even text layout of non-text lines. Currently, there are two major difficulties in widely used scene text detection methods: on the one hand, most existing methods adopt quadrilateral bounding boxes (bounding boxes), which cannot accurately locate text with arbitrary shapes; on the other hand, There are many scenes where the text lines are very similar to each other, resulting in a low detection ac...

AI Technical Summary

Problems solved by technology

Currently, there are two major difficulties in widely used scene text detection methods: on the one hand, most existing methods adopt quadrilateral bounding boxes (bounding boxes), which cannot accurately locate text with arbitrary shapes; on the other hand, There are many scenes where the text lines are very similar to each other, resulting in a low detection accuracy, and the connected text lines will be recognized as one line

Traditionally, segmentation-based methods are good at alleviating the first difficulty, but often fail to solve the second difficulty

Images