一种基于卷积神经网络的无参考立体图像质量评估算法
首发时间:2015-08-27
摘要:为了更好地对无参考图像进行质量评估,我们提出一种新的基于无参考图像的学习立体图像局部结构信息的卷积神经网络(Convolutional Neural Network, CNN)。选取立体图像对中图像块作为输入,使得网络可以有效学习到存在于图像局部结构中对人类更敏感的感知信息,并进行图像质量估计。利用多组卷积池化层,获得对结构的高等级特征表示。最后利用多层感知机(Multi-layer Perceptron, MLP)对所学习到的高层特征进行进一步学习并最终得到对立体图像质量的评估分数。针对不同的输入设计两类网络,利用2D图像的单输入网络获得网络初始值,应用于针对立体图像的双输入网络(左右视角图像)。在LIVE 3D数据库下进行测试,该无参考评估算法得到了相较于已有质量评估更好的结果。
For information in English, please click here
A no-reference stereoscopic image quality assessment based on CNN
Abstract:In this paper, to perform better on no-reference quality assessment, a new convolutional neural network (CNN) were proposed to learn the structures of stereoscopic images. Image patches from the stereoscopic images were chosen as inputs, for this reason, the proposed CNNs can learn the local structures which are sensitive to human perception and representative for perceptual quality evaluation. Following with multiple convolution and pooling layers together, the higher level representation of learned structures were composed. Multilayer perceptron (MLP) was further employed to train the learned representation to a final value to assess the perceptual quality of the stereo image. Two different networks were designed for different input, one-input CNN with 2D image patches was used to initial the weights for the two-input CNN, which with the image patches from left and right view image. With the evaluation on LIVE 3D stereoscopic image database, the proposed no-reference algorithm performed better than other quality assessment algorithms.
Keywords: Computer Neural Network No-reference Image Algorithm Stereoscopic Image Quality Assessment
论文图表:
引用
No.4652819108014914****
同行评议
勘误表
一种基于卷积神经网络的无参考立体图像质量评估算法
评论
全部评论0/1000