A method for detecting character elements in natural scene images

Abstract

The invention relates to a character element detection method in a natural scene image, belonging to the field of computers. This method first uses the maximum pooling operation to suppress the target non-maximally, then uses the feature extraction network to extract the feature, and finally integrates the top and bottom layer features through feature fusion to improve the representation ability of the feature and combine the prediction network to output the character target. exact location. By designing a novel neural network structure, the method of the invention can integrate the features of the top and bottom layers of characters in natural scene images, so as to realize accurate detection of the positions of characters arranged horizontally, obliquely and curvedly in a complex background.

Description

technical field [0001] The invention belongs to the field of computers, and relates to a character element detection method in a natural scene image. Background technique [0002] With the advancement of computer vision technology, text recognition technology has emerged and flourished rapidly. Its application has spread to all corners of the national economy and social life, and is bringing huge changes to the way of production, work and even the way of life of human beings. With the continuous deepening of applications, the market has higher and higher requirements for text detection methods in natural scenes with more complex scenes. [0003] Text detection is the prelude to subsequent text recognition and other operations, and its detection effect directly affects the effect of subsequent recognition and other operations. Text in natural scenes usually contains a lot of background interference, image lighting effects, etc., and the aspect ratio of different languages ​​...

AI Technical Summary

Problems solved by technology

Traditional detection methods are often unable to deal with many problems such as background interference. Deep learning methods can achieve better detection results, but deep learning methods rely on large data set training. The existing single-character data sets for text detection in natural scenes are due to the cost of labeling. Larger, so the data set is less, which will have a certain negative impact on algorithm development and effect testing

Images