作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程

• 人工智能及识别技术 • 上一篇    下一篇

一种简化的BP神经网络图像插值算法

钱育蓉1,2,王 谨2,郑济昌2,于 炯1,贾振红1,冷洪勇3   

  1. (1. 新疆大学软件学院,乌鲁木齐 830008;2. 汉阳大学电子和计算机工程学院,韩国 首尔 133791; 3. 新疆大学信息科学与工程学院,乌鲁木齐 830046)
  • 收稿日期:2012-08-02 出版日期:2013-09-15 发布日期:2013-09-13
  • 作者简介:钱育蓉(1980-),女,副教授,主研方向:图像压缩,人工智能;王 谨,博士研究生;郑济昌、于 炯(通讯作者)、 贾政红,教授、博士生导师;冷洪勇,硕士研究生
  • 基金资助:
    国家自然科学基金资助项目(61262088, 61063042, 61363083);新疆自然科学基金资助项目(2011211A011, 2013211A011);新疆高校科研计划基金资助项目(XJEDU2012I10);新疆大学博士启动基金资助项目(BS100128)

A Simplified Image De-interlacing Algorithm in Back Propagation Neural Network

QIAN Yu-rong 1,2, WANG Jin 2, JEONG Je-chang 2, YU Jiong 1, JIA Zhen-hong 1, LENG Hong-yong 3   

  1. (1. School of Software, Xinjiang University, Urumqi 830008, China; 2. School of Electronics & Computer Engineering, Hanyang University, Seoul 133791, Korea; 3. School of Information Science and Engineering, Xinjiang University, Urumqi 830046, China)
  • Received:2012-08-02 Online:2013-09-15 Published:2013-09-13

摘要: 为进一步提高图像插值质量,丰富非线性图像插值算法研究,提出一种简化的神经网络插值算法。利用前向反馈神经网络(BP-NN)构造最佳的图像插值模型,通过2组实验确定该模型的BP网络最佳拓扑结构、最佳采样模型和采样点数量,并定量描述每组模型的耗时。实验结果表明,对512×512像素图像采用BP-NN训练点数量为50 000、拓扑结构为8-16-1的参数插值时,该算法平均插值时间约为0.7 s,且其峰值信噪比比线性均值方法平均高1 dB~2 dB,能够得到更佳的视觉感受。

关键词: 前向反馈神经网络, 图像插值, 峰值信噪比, 采点模式, 隐藏层神经元, 线性插值

Abstract: The objective of this paper is to improve the quality of image interpolation, enrich non-linear image interpolation algorithm. This paper presents a simplified neural network interpolation algorithm, in which the Back Propagation Neural Network(BP-NN) is adopted to construct the best interpolation model. Two sets of experiments determine the best topology of the BP network’s model, the optimal sampling model, the number of sampling points, and describe the time-consuming of each group model quantitatively. Experimental results show that a 512×512 pixel image interpolation using BP-NN can obtain the Peak Signal to Noise Radio(PSNR) 1dB~2 dB higher than Linear Average(LA) method, while the number of pixels in training sets is 50 000 with the topology of 8-16-1. Therefore, the proposed algorithm performs better visual quality.

Key words: Back Propagation Neural Network(BP-NN, image de-interlacing, Peak Signal to Noise Radio(PSNR), sampling mode, hidden layer neuron, linear interpolation

中图分类号: