摘要
为了提高图像插值质量,利用前向反馈人工神经网络(BP-ANN)的自学习、自适应和泛化能力,开展分区BPANN图像差值研究.将图像中待插值像素划分为光滑区和边缘区,每个区分别对应1个BP-ANN进行图像差值操作,并通过3组实验确定分区BP-ANN的网络结构、采点模式和插值流程.结果表明,采用8-16-1拓扑结构的BP-ANN算法可达到图像的可视化质量和时间之间的最佳平衡点;与经典线性均值(LA)图像插值算法相比,分区BP-ANN算法在保持最佳视觉效果的前提下,峰值性噪比高约0.593 8dB.
In order to improve the quality of image interpolation,back-propagation artificial neural network（BP-ANN）with self-learning,adaptive and generalization ability has been used to carry out the regional image interpolation research.Missing pixel in image is divided into smooth region and edge region,each region corresponding to a BP-ANN interpolation operation.To identify the topology structure,sampling mode and the interpolation process of regional BP-ANN,three experiments were performed.The experimental results show that best balance between CPU-time consumption and visual quality lies in 8-16-1topology of BP-ANN;compared with the classical LA interpolation algorithm,our proposed algorithm provides 0.5938dB higher PSNR while performed better visual quality.
出处
《吉首大学学报（自然科学版）》
CAS
2016年第3期28-33,共6页
Journal of Jishou University(Natural Sciences Edition)
基金
国家自然科学基金资助项目(61462079
61363083
61262088)
新疆维吾尔自治区青年博士科技人才创新项目(2013731004)
新疆维吾尔自治区高校科研计划项目(XJEDU2012I10)
新疆维吾尔自治区自然科学基金资助项目(2013211A011
2011211A011)
关键词
插值
前向反馈人工神经网络
线性均值插值
峰值性噪比
网络拓扑
interpolation
back propagation artificial neural network
linear average interpolation
peak signal to noise radio
network topology
