基于机器学习的视频中的数字时钟识读的研究

发布时间：2018-05-19 04:18

本文选题：机器学习 + 视频中数字时钟定位　；参考：《华中师范大学》2017年硕士论文

【摘要】：识读视频中的数字时钟是一个重要的应用问题,在智能监控、娱乐等领域有着重要的应用价值。文献中已有算法可以很好地识读视频中的数字时钟,但是这样的算法还存在两方面的问题。其一是这些算法以启发式方法为主,没有利用强有力的机器学习方法。其二是这些算法有一批参数需要手动调整。针对这两方面的问题,本文提出全面使用机器学习方法完成时钟的定位和时钟数字识别。本文是首次研究无识别的机器学习定位方法。其思想如下:与传统方法不同的是,在定位阶段,先根据秒像素点灰度值变化的周期性快速定位到秒像素的数字位置,且首次引入了机器学习的思想解决此问题,更具有智能性。其次,考虑了数字时钟通常有稳定的背景,及时钟不同数字的前景、背景的颜色不同,快速定位到时钟的时、分区域,先利用机器学习区分颜色,进而采用机器学习进行图像分割,但因为利用了颜色信息,效率更高。时钟识读阶段,本研究主要采用了卷积神经网络,在单个数字识别上,取得了比CRF更高的效率,且降低了对字体等因素的依赖。因秒数字具有周期性的系列特点,本研究也系列识别上作了一定探讨,用来矫正单个数字识别的错误。最后,本研究也在如何借鉴先前的启发式函数的成果经验,如何用机器学习的方法针对具体问题设计适合的网络结构,及模拟启发式函数上,做了一定探讨。
[Abstract]:Digital clock in video reading is an important application problem, which has important application value in intelligent monitoring, entertainment and other fields. The existing algorithms in the literature can read the digital clock in video very well, but there are still two problems in this algorithm. One is that these algorithms are mainly heuristic, without powerful machine learning methods. The second is that these algorithms have a number of parameters to be manually adjusted. In order to solve these two problems, this paper proposes a machine learning method to complete clock localization and clock digital recognition. This paper is the first time to study the localization method of unrecognized machine learning. The idea is as follows: different from the traditional method, in the localization stage, the digital position of the second pixel is quickly located according to the periodicity of the change of the gray value of the second pixel, and the idea of machine learning is introduced to solve this problem for the first time. More intelligent. Secondly, considering the stable background of digital clock and the foreground of different digital clock, the color of background is different. When the clock is located quickly, it is divided into regions, and the color is distinguished by machine learning. Then machine learning is used for image segmentation, but because of the use of color information, the efficiency is higher. In the phase of clock recognition, convolution neural network is used in this study, which achieves higher efficiency than CRF in single digit recognition, and reduces the dependence on font and other factors. Because the second digit has the periodic series characteristic, this research also makes the certain discussion on the series recognition, which is used to correct the error of the single digit recognition. Finally, this study also discusses how to draw lessons from the previous experience of heuristic function, how to design a suitable network structure for specific problems by using machine learning methods, and how to simulate heuristic function.
【学位授予单位】：华中师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP181

【相似文献】