Natural scene image text detection method based on cascaded convolutional neural network

Abstract

The invention discloses a natural scene image text detection method based on a cascaded convolutional neural network, which comprises the steps of firstly performing candidate character extraction onan input image by using a maximally stable extremal region method; secondly, connecting a pre-trained 16-net convolutional neural network and 32-net convolutional neural network in series to form a cascaded convolutional neural network, classifying the candidate characters by using the cascaded convolutional neural network, eliminating inauthentic characters in the candidate characters to obtain afinal authentic character detection result; then pairing every two adjacent characters in the character detection result, combining the characters having a shared character until all of the characters are combined, and obtaining a text line; and finally dividing the text line into a plurality of words or phrases according to the character spacing, and acquiring a final text detection result. Thenatural scene image text detection method effectively solves a problem that training samples are unbalance in category, and effectively improves the text detection efficiency and performance.

Description

technical field [0001] The invention belongs to the technical field of natural scene text detection, and in particular relates to a natural scene image text detection method based on a cascaded convolutional neural network. Background technique [0002] Natural scene text detection generally includes four steps: candidate character extraction, character and non-character classification, text line construction and text line segmentation. Candidate character extraction refers to extracting regions with such features from natural scene images as candidate characters by studying the structure and color features of characters; then using machine learning algorithms or deep learning algorithms to train a binary classifier, and using this The classifier classifies the candidate characters and divides them into characters and non-characters; then uses the positional relationship between adjacent characters to group the characters into text lines, that is, the text line structure; fi...

AI Technical Summary

Problems solved by technology

The problem with the penalty factor at the algorithm level is how to set the penalty factor and how to determine the size of the penalty factor, and when the sample categories are extremely unbalanced, this method will fail

Images