Text detection method in natural scene based on deep learning

Abstract

The invention provides a text detection method in a natural scene based on deep learning, which belongs to the field of computer vision. The text detection method is specifically composed of a starting module, a character area recognition network and a capsule screening and classifying network. The method comprisesthe steps of firstly, enabling a starting module to reduce the parameters of a network, and placing convolution kernels of different sizes in the same convolution layer, so that the adaptability of the network to character features of different scales is enhanced; then detecting single characters in the picture by using a character region recognition network, and screening and classifying the detected single characters by using a capsule classification network so as to judge whether the detected single characters are characters; and finally, splicing the single characters into a text line through a character splicing method based on multi-feature map fusion and hole convolution. Compared with the prior art, the influence of factors such as the direction, the size and the position of the characters on the detection result can be reduced, so that the correct rate of character detection is ensured, and the robustness and the accuracy are further improved.

Description

technical field [0001] The invention relates to the technical field of image processing, in particular to a deep learning-based text detection method in natural scenes. Background technique [0002] With the continuous development and popularization of computer technology and smart devices, there are many image data in complex scenes, such as street view advertising pictures, handwritten forms, etc. They are distributed in every corner of the Internet, and the information is very rich and valuable. Therefore, it has become an unavoidable demand to use computers to process the increasing variety of digital image data. How to extract the required and useful information from these images is becoming more and more important, which is also a hot research field in the field of computer vision. [0003] In these digital images, text information is the most direct information. Understanding text information in natural scene images has many practical significance for human-computer ...

AI Technical Summary

Problems solved by technology

However, there are many existing text detection algorithms based on deep learning. When performing text detection on complex background images, the text will be detected as a target object, because the aspect ratio (also called aspect ratio, w / h) of the text detection object The change is large, unlike the relatively small change in the aspect ratio of the object. In addition, because the text in the natural scene is affected by factors such as size and position, it will add a lot of difficulties to the detection

Images