论文部分内容阅读
极限学习机(Extreme Learning Machine,ELM)较其它神经网络具有训练速度快、泛化能力强的特点.然而对于高维的立体图像数据,无论ELM还是传统神经网络均需经过特征提取的预处理,但是传统特征提取的方式是否真正符合人的感知特性有待进一步研究.深度学习是一种模拟人脑深层次学习的神经网络,因此提出基于深度结构的极限学习机算法(Deep Extreme Learning M achine,D-ELM),该方法通过深度学习预训练来逐层表达输入数据的分布式特征,从而实现原始数据的特征提取.实验结果表明,深度结构下的ELM网络更加稳定高效,对于250幅不同等级的立体图像样本进行测试后的准确率达到了96.11%.此外,本文还分析了隐节点数对网络的影响,而且将D-ELM与ELM、支持向量机等在立体图像质量评价上的性能进行了比较.
Extreme Learning Machine (ELM) is faster and more generalized than other neural networks.However, for high-dimensional stereoscopic image data, both ELM and traditional neural networks need to be pretreated by feature extraction, However, whether the traditional feature extraction method is in line with the human perception characteristics remains to be further studied.Draw learning is a kind of neural network that simulates human brain deep learning, so Deep Extreme Learning Algorithm (D -ELM), this method through the depth of learning pre-training to express the distributed features of the input data layer by layer, so as to achieve the original data feature extraction.Experimental results show that the depth structure of the ELM network is more stable and efficient for 250 different levels The accuracy of stereo image samples tested is 96.11% .In addition, the paper also analyzes the impact of hidden nodes on the network, and the performance of D-ELM, ELM, support vector machines and other three-dimensional image quality evaluation Compare