基于神经网络的立体图像质量客观评价

发布时间：2018-04-21 13:28

本文选题：立体图像 + 客观评价　；参考：《天津大学》2016年硕士论文

【摘要】：随着立体成像技术的不断发展,准确、有效地评价立体图像质量已成为立体技术领域的研究热点及难点之一。立体图像质量的评价方法分为主观评价和客观评价两种。主观评价由合格被试依据自身主观感受对测试图像质量给出评分,这种方法能够真实准确地反映图像的质量,但它耗时耗力,且可操作性较差。因而,建立一套有效的立体图像质量客观评价模型已成为立体成像技术的重点研究课题之一。论文在对立体图像质量评价的研究背景、发展现状、发展趋势及其他相关理论进行阐述的基础上,考虑到目前人类视觉系统的相关研究仍存在较大的局限性,提出采用正交局部保留投影和极端学习机的方法建立立体图像质量评价系统。鉴于立体图像具有复杂度高、信息量大的特点,论文选取正交局部保留投影法对图像进行有效地降维处理,该方法可以在对图像降维的同时保留不同类别图像间的结构,可以更有效地提取出立体图像的特征。极端学习机网络具有参数选择简单、泛化性好等特点,但是该网络具有一定的随机性。鉴于此,论文提出采用经过遗传算法优化的极端学习机作为分类器,使评价系统可以获取更好的分类识别性能。本文选取了380幅经过不同失真处理、覆盖不同评分等级的立体图像,其中154幅为训练样本,226幅为测试样本。实验结果表明,选用正交局部保留投影法作为特征提取方法,使用ELM分类器在测试样本中的客观评分正确率可以达到93.36%,比选用主成分分析法所能达到的92.03%的准确率有更好的表现。使用遗传算法对网络参数进行优化后,ELM网络的分类正确率可以达到96.03%,使评价系统的准确率有了明显的提高。此外,本文还对不同神经网络分类器的质量评价性能进行了分析比较。
[Abstract]:With the development of stereo imaging technology, accurate and effective evaluation of stereo image quality has become one of the hot and difficult research fields. The evaluation methods of stereo image quality can be divided into subjective evaluation and objective evaluation. The subjective evaluation is evaluated by the qualified subjects according to their subjective feelings. This method can reflect the image quality truthfully and accurately, but it is time-consuming and energy consuming, and its operability is poor. Therefore, the establishment of an effective objective evaluation model of stereo image quality has become one of the key research topics of stereo imaging technology. Based on the research background, development status, development trend and other related theories of stereo image quality evaluation, this paper considers that there are still some limitations in the research of human visual system. A stereo image quality evaluation system based on orthogonal local preserving projection and extreme learning machine is proposed. In view of the high complexity and large amount of information of the stereo image, the orthogonal local preserving projection method is selected to reduce the dimension of the image effectively. This method can reduce the dimension of the image while preserving the structure of different kinds of images. The feature of stereo image can be extracted more effectively. Extreme learning machine network is characterized by simple parameter selection, good generalization and so on, but it has some randomness. In view of this, the paper proposes to use the extreme learning machine optimized by genetic algorithm as the classifier, so that the evaluation system can obtain better classification and recognition performance. In this paper, 380 stereo images with different distortion processing are selected, of which 154 are training samples and 226 are test samples. The experimental results show that the orthogonal locally reserved projection method is used as the feature extraction method. The objective scoring accuracy of ELM classifier in test samples can reach 93.36, which is better than the 92.03% accuracy of principal component analysis. After the optimization of network parameters by genetic algorithm, the classification accuracy of ELM network can reach 96.03, which improves the accuracy of the evaluation system obviously. In addition, the quality evaluation performance of different neural network classifiers is analyzed and compared.
【学位授予单位】：天津大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41;TP183

【相似文献】