查询结果:   郁松,彭志文.基于卷积神经网络的自然背景字符识别[J].计算机应用与软件,2017,34(12):228 - 234.
中文标题
基于卷积神经网络的自然背景字符识别
发表栏目
图像处理与应用
摘要点击数
608
英文标题
NATURAL BACKGROUND CHARACTER RECOGNITION BASED ON CONVOLUTIONAL NEURAL NETWORK
作 者
郁松 彭志文 Yu Song Peng Zhiwen
作者单位
中南大学软件学院 湖南 长沙 410075     
英文单位
School of Software,Central South University,Changsha 410075,Hunan,China     
关键词
自然背景字符识别 卷积神经网络 图像分类
Keywords
Recognition of scene characters Convolutional neural network Image classification
基金项目
作者资料
郁松,副教授,主研领域:图像处理,数据挖掘。彭志文,硕士。 。
文章摘要
随着计算机视觉技术的发展,自然背景中字符的识别在图片检索、视频检索、无人车识别周围场景信息等领域都扮演了不可或缺的角色。相对于手写字符、打印字符的识别,自然背景字符的识别有着光照强度变化大、背景纹理复杂、字体样式和颜色多变等特点,这都给识别带来了巨大的挑战。主要是基于LeNet-5的网络结构设计了一种适合于识别自然背景字符的卷积神经网络,由于在这一领域以往的研究工作的基准数据集是较小的数据集(Chars74K-15),为了便于比较,实验也是基于同样的数据集。但因为卷积神经网络是在巨大数据量的驱动下才会有良好的效果,因此还提出了一种预处理方式和fine-tune相结合用于解决自然背景字符图片数据量较小的问题。
Abstract
With the development of the computer vision technology, the recognition of characters in natural background plays an indispensable role in the fields of picture retrieval, video retrieval and unmanned vehicle recognition. Compared to the recognition of handwritten characters and printed characters, the natural scene characters have many different features. For example, the variation of light intensity, complex background texture, the variation of font’s style and color. All these features bring a huge challenge to the recognition. The paper raised a CNN which can recognize natural scene characters effectively. Most of the past research is based on Chars74K-15 which does not contain many images. In order to compare with the past, we used the same data set. Because of the large amount of data on training the CNN, we raised a preprocessing method with fine-tune to solve the problem of lacking data.
下载PDF全文