Text target detection method and system based on a deep network

Abstract

The invention discloses a text target detection method and system based on a deep network, and the method comprises the steps: selecting an original image, and extracting a feature map of the originalimage; judging whether the anchor point frame of the feature map is a foreground or a background, and correcting the anchor point frame by using a frame regression device to obtain a proposal area; collecting the input feature map and the proposal area, and extracting the feature map of the proposal area; segmenting the feature map of the proposal area into a left upper corner point feature map,a right upper corner point feature map, a right lower corner point feature map and a left lower corner point feature map of the text target; and performing regression of corresponding angular point coordinates on each feature map to obtain a frame of the original picture.

Description

technical field [0001] The invention relates to the technical field of text detection, in particular to a text object detection method and system based on a deep network. Background technique [0002] Due to the lack of obvious edge features of documents and bills, it is difficult to summarize image features, and traditional machine learning methods cannot achieve accurate positioning. Based on the method of deep learning, a large amount of data suitable for actual application scenarios is obtained through data collection, data augmentation, etc., and then the neural network is used to automatically learn useful features, avoiding the deviation of artificially defined features, so as to ensure the learned features. are the most useful features for final text object detection. [0003] Faster Rcnn is a target detection framework proposed in 2016, and it is still one of the mainstream target detection frameworks until now. Structurally, Faster Rcnn has integrated feature ext...

AI Technical Summary

Problems solved by technology

[0005] The target positions detected by the conventional target detection framework are all rectangles, and if the target rotation angle is large or there is a certain transparency, there is a big difference between the detected frame and the actual target position, which is necessary for the use of these For the process of detecting the location for subsequent processing, it is tantamount to adding a lot of uncertainty;

Images