應(yīng)用深度極限學(xué)習(xí)機(jī)的立體圖像質(zhì)量評(píng)價(jià)方法
發(fā)布時(shí)間：2018-10-10 12:46
【摘要】：極限學(xué)習(xí)機(jī)(Extreme Learning Machine,ELM)較其它神經(jīng)網(wǎng)絡(luò)具有訓(xùn)練速度快、泛化能力強(qiáng)的特點(diǎn).然而對(duì)于高維的立體圖像數(shù)據(jù),無論ELM還是傳統(tǒng)神經(jīng)網(wǎng)絡(luò)均需經(jīng)過特征提取的預(yù)處理,但是傳統(tǒng)特征提取的方式是否真正符合人的感知特性有待進(jìn)一步研究.深度學(xué)習(xí)是一種模擬人腦深層次學(xué)習(xí)的神經(jīng)網(wǎng)絡(luò),因此提出基于深度結(jié)構(gòu)的極限學(xué)習(xí)機(jī)算法(Deep Extreme Learning M achine,D-ELM),該方法通過深度學(xué)習(xí)預(yù)訓(xùn)練來逐層表達(dá)輸入數(shù)據(jù)的分布式特征,從而實(shí)現(xiàn)原始數(shù)據(jù)的特征提取.實(shí)驗(yàn)結(jié)果表明,深度結(jié)構(gòu)下的ELM網(wǎng)絡(luò)更加穩(wěn)定高效,對(duì)于250幅不同等級(jí)的立體圖像樣本進(jìn)行測試后的準(zhǔn)確率達(dá)到了96.11%.此外,本文還分析了隱節(jié)點(diǎn)數(shù)對(duì)網(wǎng)絡(luò)的影響,而且將D-ELM與ELM、支持向量機(jī)等在立體圖像質(zhì)量評(píng)價(jià)上的性能進(jìn)行了比較.
[Abstract]:Extreme Learning Machine (Extreme Learning Machine,ELM) has the advantages of faster training speed and better generalization ability than other neural networks. However, for high-dimensional stereo image data, both ELM and traditional neural networks need to be preprocessed by feature extraction, but whether the traditional feature extraction method really accords with human perception needs further study. Deep learning is a neural network that simulates the deep learning of human brain. Therefore, an algorithm of extreme learning machine (Deep Extreme Learning M achine,D-ELM) based on depth structure is proposed, which can express the distributed feature of input data layer by layer through the pre-training of depth learning. In order to achieve the feature extraction of the original data. The experimental results show that the ELM network with depth structure is more stable and efficient, and the accuracy of testing 250 stereo images of different grades is 96.11. In addition, the influence of the number of hidden nodes on the network is analyzed, and the performance of D-ELM and ELM, support vector machine in stereo image quality evaluation is compared.
【作者單位】： 天津大學(xué)電氣自動(dòng)化與信息工程學(xué)院;
【基金】：國家自然科學(xué)基金項(xiàng)目(61002028)資助 國家“八六三”計(jì)劃項(xiàng)目(2012AA011505,2012AA03A301)資助
【分類號(hào)】：TP18;TP391.41
[Abstract]:Extreme Learning Machine (Extreme Learning Machine,ELM) has the advantages of faster training speed and better generalization ability than other neural networks. However, for high-dimensional stereo image data, both ELM and traditional neural networks need to be preprocessed by feature extraction, but whether the traditional feature extraction method really accords with human perception needs further study. Deep learning is a neural network that simulates the deep learning of human brain. Therefore, an algorithm of extreme learning machine (Deep Extreme Learning M achine,D-ELM) based on depth structure is proposed, which can express the distributed feature of input data layer by layer through the pre-training of depth learning. In order to achieve the feature extraction of the original data. The experimental results show that the ELM network with depth structure is more stable and efficient, and the accuracy of testing 250 stereo images of different grades is 96.11. In addition, the influence of the number of hidden nodes on the network is analyzed, and the performance of D-ELM and ELM, support vector machine in stereo image quality evaluation is compared.
【作者單位】： 天津大學(xué)電氣自動(dòng)化與信息工程學(xué)院;
【基金】：國家自然科學(xué)基金項(xiàng)目(61002028)資助 國家“八六三”計(jì)劃項(xiàng)目(2012AA011505,2012AA03A301)資助
【分類號(hào)】：TP18;TP391.41
【相似文獻(xiàn)】
相關(guān)期刊論文 前10條
1 韓偉;;日本立體圖像技術(shù)近十年的回顧與前瞻[J];有線電視技術(shù);2011年07期
2 楊s，
本文編號(hào)：2261816
本文鏈接：http://sikaile.net/kejilunwen/zidonghuakongzhilunwen/2261816.html
最近更新
教材專著
