新型工业化
新型工業化
신형공업화
New Industrialization Straregy
2012年
2期
33-45
,共13页
潘瑜%孙权森%郑钰辉%孙怀江%夏德深
潘瑜%孫權森%鄭鈺輝%孫懷江%夏德深
반유%손권삼%정옥휘%손부강%하덕심
模式识别%核广义典型相关分析%广义典型相关分析%维纳滤波
模式識彆%覈廣義典型相關分析%廣義典型相關分析%維納濾波
모식식별%핵엄의전형상관분석%엄의전형상관분석%유납려파
pattern recognition%kernel generalized canonical correlation analysis%generalized canonical correlation analysis%wiener filter
为了提高基于学习的维纳滤波方法性能,将广义典型相关分析理论推广到核空间中,并在核空间中将此方法与维纳滤波相结合,提出了基于核广义典型相关分析的维纳滤波及其快速算法。该方法首先使用主成分分析对噪声图像进行预处理,而后将处理后图像数据映射到高维核空间中,使用核技巧依据核广义典型相关分析理论抽取相关特征来计算维纳滤波所需的降秩估计量,最后利用维纳滤波的均方误差最小的理念获取线性空间内的图像恢复结果。为了减少在特征抽取过程中的计算量,以空间变换的方法减少了矩阵维数;为了进一步提升图像恢复效果,在维纳滤波中引入了保真项。实验表明,该方法所抽取的相关特征能够降低图像恢复结果的错误率,并且该恢复过程对于降秩估计量秩的大小以及算法迭代次数具有健壮性,快速算法能够在保持图像质量的情况下减少25%以上的时间消耗。
為瞭提高基于學習的維納濾波方法性能,將廣義典型相關分析理論推廣到覈空間中,併在覈空間中將此方法與維納濾波相結閤,提齣瞭基于覈廣義典型相關分析的維納濾波及其快速算法。該方法首先使用主成分分析對譟聲圖像進行預處理,而後將處理後圖像數據映射到高維覈空間中,使用覈技巧依據覈廣義典型相關分析理論抽取相關特徵來計算維納濾波所需的降秩估計量,最後利用維納濾波的均方誤差最小的理唸穫取線性空間內的圖像恢複結果。為瞭減少在特徵抽取過程中的計算量,以空間變換的方法減少瞭矩陣維數;為瞭進一步提升圖像恢複效果,在維納濾波中引入瞭保真項。實驗錶明,該方法所抽取的相關特徵能夠降低圖像恢複結果的錯誤率,併且該恢複過程對于降秩估計量秩的大小以及算法迭代次數具有健壯性,快速算法能夠在保持圖像質量的情況下減少25%以上的時間消耗。
위료제고기우학습적유납려파방법성능,장엄의전형상관분석이론추엄도핵공간중,병재핵공간중장차방법여유납려파상결합,제출료기우핵엄의전형상관분석적유납려파급기쾌속산법。해방법수선사용주성분분석대조성도상진행예처리,이후장처리후도상수거영사도고유핵공간중,사용핵기교의거핵엄의전형상관분석이론추취상관특정래계산유납려파소수적강질고계량,최후이용유납려파적균방오차최소적이념획취선성공간내적도상회복결과。위료감소재특정추취과정중적계산량,이공간변환적방법감소료구진유수;위료진일보제승도상회복효과,재유납려파중인입료보진항。실험표명,해방법소추취적상관특정능구강저도상회복결과적착오솔,병차해회복과정대우강질고계량질적대소이급산법질대차수구유건장성,쾌속산법능구재보지도상질량적정황하감소25%이상적시간소모。
To improve the performance of wiener filter based on learning, this paper extends the Generalized Canonical Correlation Analysis (GCCA) from the linear space to the nonlinear kernel space and addresses the kernel wiener filter based on kernel GCCA and its corresponding fast algorithm. To acquire the restored images, the Principal Components Analysis is firstly used to process the noisy images. Secondly, the processed image data sets are transformed to the high dimensional space. Then, with the kernel trick and the features extracted by Kernel Generalized Canonical Correlation Analysis, the reduced rank estimator needed by kernel wiener filter is calculated. After that, the kernel wiener filter which solves problems with the method of minimizing the mean square error is used to acquire the restored images in the original space. To reduce the computation during the feature extraction, a space transform method is used to reduce the dimension of matrices. To improve the effect of the restored image, the fidelity term is added into the wiener filter. The experiment demonstrates that the features extracted by the new method are able to reduce the error rate of the restored images and the new method is robust to the rank of the reduced rank estimator and the iterative times of the arithmetic. Additionally, the fast algorithm can reduce the time consuming at least 25% while preserving the quality of restored image.