Scene text detection method based on corner recognition

A text detection and corner point technology, which is applied in the field of scene text detection based on corner point recognition, can solve the problems of poor accuracy and efficiency of text detection, inability to effectively generate bounding boxes, and inability to effectively complete text detection, etc., to reduce calculation Quantity, the effect of ensuring accuracy and efficiency

CN114038002APending Publication Date: 2022-02-11七腾机器人有限公司
1 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Current Assignee / Owner
Publication Date
2022-02-11

Smart Images

  • Figure 1
    Figure 1
  • Figure 2
    Figure 2
  • Figure 3
    Figure 3
Patent Text Reader

Abstract

The invention relates to the technical field of internet information detection and image text recognition, in particular to a scene text detection method based on corner recognition, which comprises the following steps: acquiring an image to be detected; extracting image features of the to-be-detected image; predicting angular point information of each angular point in the corresponding orientation based on the image features of the to-be-detected image, where the angular point information comprises an angular point thermodynamic diagram used for positioning angular point coordinates and an embedded vector used for combining angular points; positioning the angular point coordinates of each angular point in the corresponding orientation based on the angular point thermodynamic diagram, then identifying each angular point which belongs to the same text instance and can form a corresponding bounding box based on an embedded vector, and constructing an angular point combination based on the corresponding angular point; and generating a corresponding bounding box based on the angular point coordinates in the angular point combination. According to the scene text detection method based on corner recognition, the accuracy and efficiency of text detection can be ensured.
Need to check novelty before this filing date? Find Prior Art

Description

technical field

[0001] The invention relates to the technical field of Internet information detection and image text recognition, in particular to a scene text detection method based on corner point recognition. Background technique

[0002] With the rise and wide application of Internet information technology and the emergence of various handheld devices equipped with cameras, the application of image information detection technology based on natural scenes is increasing. In the era of big data information, for various targets in natural scenes The detection of information such as individuals, texts, and data is becoming more and more important, and text detection and recognition in natural scenes has new developments. Text detection and recognition in natural scenes consists of two steps: text detection and text recognition. The task of text detection is to obtain the bounding box (Bounding Boxes) of the smallest text area contained in the text image. Good text detection ...

Examples

Embodiment

[0059] This embodiment discloses a scene text detection method based on corner recognition.

[0060] Such as figure 1 and figure 2 As shown, the scene text detection method based on corner recognition includes the following steps:

[0061] S1: Acquire the image to be detected;

[0062] S2: Extract the image features of the image to be detected; use two stacked Hourglass networks as the backbone network to extract multi-scale image features (multi-scale image features can improve the accuracy of text detection); the two stacked Hourglass networks are connected through a ReLU module and a residual module composed of 256 channels;

[0063] For a single Hourglass network, before the image is input into the Hourglass network, a 7×7 convolution module with a step size of 2 and a number of convolution kernels of 128, and a residual structure with a step size of 2 and a number of convolution kernels of 256 will be The resolution of the image is reduced to 1 / 4 of the original reso...