Text detection and recognition method based on convolutional neural network and in natural scene

Abstract

The invention provides a text detection and recognition method based on a convolutional neural network and in a natural scene. The method includes: training a quick area convolutional neural network to acquire a network for text detection in the natural scene to perform text detection in the natural scene; collecting vocabularies to acquire a dictionary; learning features of text vocabularies to train a convolutional neural network classifier, and classifying input areas to realize text recognition. By utilizing the method, good text detection and recognition results can be acquired under the background of different complex degrees and different text fonts and colors.

Description

technical field [0001] The invention belongs to the technical field of computer vision, object detection and recognition, and specifically relates to a text detection and recognition method in a natural scene based on a fast regional convolutional neural network and a convolutional neural network. Background technique [0002] Compared with the visual cues of traditional low-level features, text rich in clear high-level semantic features provides useful information in many visual tasks, so detecting and recognizing text in natural scenes is an important technology in computer vision . Usually, due to complex natural scenes, variable text distribution, fonts, colors, uneven lighting effects, and camera resolution, it is difficult to detect and recognize text in natural scenes. Currently, there are two types of text detection and recognition methods: [0003] The first is a method with a clear distinction between text detection and recognition. The method generally includes...

AI Technical Summary

Problems solved by technology

This type of method has an obvious distinction between the text detection stage and the recognition stage. The advantages are: the overall process is from coarse to fine, and many background areas are filtered out during the detection and removal of non-text areas, which greatly reduces the The calculation cost is reduced; the disadvantage is that when irrelevant methods are used in each stage, the overall complexity of this type of method is very high, and the optimization of parameters in each stage will become more difficult, which will easily lead to error propagation

The advantage is that it can avoid difficult character segmentation, but the disadvantage is that the calculation cost is relatively high

Images