Method for detecting scene characters based on fragments and links of convolutional neural network

Abstract

The invention discloses a method for detecting scene characters based on fragments and links of a convolutional neural network. A fragment link model constructed in the invention comprises a pluralityof convolution feature layers and convolution predictors which are connected in sequence, and detection efficiency is very high because fragments and links can be intensively detected on multiple scales at the same time in the forward conduction process. The link types are specifically divided into same-layer links and cross-layer links, the same-layer links are connected with segments detectedin the same feature layer, and the cross-layer links can be connected with segments on different layers. Fragments on the same or different scales can be combined by using the cross-layer link and thesame-layer link, so that the problems of missing inspection and repeated inspection are well avoided.

Description

technical field [0001] The invention belongs to the application field of digital image processing, and in particular relates to a method for detecting scene characters based on convolutional neural network segments and links. Background technique [0002] Understanding images is a major goal of computer vision. The understanding of images is divided into different levels. For example, the edge detection of objects is the underlying image understanding; the semantic segmentation of objects is the middle level of understanding and so on. The understanding of the text carried in the image is the understanding of high-level semantics. This information is compatible with the human symbol system and can be directly used for high-level semantic and logical analysis. Due to the ubiquity of text and the importance of text information, understanding text in pictures has always been in an important position in computer vision. The technology of recognizing text from images is usually...

AI Technical Summary

Problems solved by technology

However, due to the difference in shape between scene text and general objects, the general object detection method is not an ideal solution.

First of all, the general object detection method is limited by its candidate region extraction algorithm, and the aspect ratio of the output bounding box can only be changed in a small range.

Therefore, it is difficult to be used to detect non-Latin characters such as Chinese and Japanese, because there is no space between words in these characters, and the detection target is often a slender bounding box with an extreme aspect ratio; secondly, the general object detection method can only Output horizontal bounding box, cannot detect non-horizontal text

Images