Broadcast and television caption recognition based automatic training data generation and deep learning method

Abstract

The invention discloses a broadcast and television caption recognition based automatic training data generation and deep learning method. The method includes the steps: S1, simulating broadcast and television caption data, and constructing a mass video character training set; S2, extracting statistic character features in the mass video character training set, and quantifying the simulated caption data; S3, using a deep neural network to train a caption recognition model, and capturing a topological structure of a caption text in broadcast video; S4, realizing individual character recognition and output of the caption text in the broadcast video according to the caption recognition model obtained by training. According to the method, by automatically constructing mass simulation training data, the problem that mass tagged data are difficult to acquire is effectively solved; by combination of a novel deep learning method and a conventional character recognition technique, the topological structure of the caption text in the broadcast video is captured accurately, so that the problem that an existing character recognition technique is unsuitable for recognizing broadcast and television caption texts is solved.

Description

technical field [0001] The invention relates to the technical field of new media content management and distribution of radio and television oriented to the integration of three networks, in particular to an automatic generation of training data and a deep learning method based on subtitle recognition of radio and television. Background technique [0002] With the continuous development of information technology and communication technology, a large amount of broadcast video information (all kinds of news, TV programs, Internet TV, etc.) is emerging, and broadcast video has gradually become an important medium for people to obtain daily information. According to the data released by the National Bureau of Statistics in 2011, as of 2011, the comprehensive population coverage rate of radio and television programs in my country has reached 97.6%. Huge social benefits and commercial value. [0003] Subtitle text in broadcast video is a kind of high-level semantic information, wh...

AI Technical Summary

Problems solved by technology

[0005] 1. Subtitle text in broadcast video is low resolution

The resolution of text characters in traditional scanned documents is generally not lower than 300dpi, and the background is extremely simple; while the height of subtitle text in broadcast videos is often only a dozen to thirty pixels, and the background is extremely complex, which leads to traditional character recognition. The technical performance plummeted and the effect was unacceptable

[0006] 2. The background of the subtitle text in the broadcast video is complex and the subtitle effects are diverse

However, due to the existence of massive complex and diverse video resources and advanced non-linear editing tool software, the subtitle texts in a large number of broadcast videos often have extremely complex backgrounds, and the artistic effects of subtitle texts are rich and colorful, which directly leads to the degradation of image binarization performance, and further The final effect of limiting character recognition

Images