K均值聚类和支持向量数据描述的图像超分辨率算法

发布时间：2019-01-30 20:04

【摘要】：目的为了提高图像超分辨率算法对数据奇异点的鲁棒性,提出一种采用K均值聚类和支持向量数据描述的图像超分辨率重建算法(Kmeans-SVDD)。方法训练过程:首先用K均值聚类算法将训练图像的近似子带划分为若干类,然后用支持向量数据描述去除每类数据的奇异点,最后在小波域内用主成分分析训练近似子带和细节子带字典。测试过程:根据同一场景高低分辨率图像近似子带相似这一现象,首先将待重建低分辨率测试图像的近似子带作为相应高分辨率测试图像的近似子带,然后由训练得到的字典恢复出高分辨率测试图像的细节子带,最后通过逆小波变换得到高分辨率测试图像。结果相比于当前双三次插值、Zeyde、ANR与Kmeans-PCA算法,Kmeans-SVDD算法重建的高分辨率测试图像的平均峰值信噪比依次提高了1.82 dB、0.37 dB、0.30 dB、0.15 dB。结论通过大量实验发现,在字典训练之前加入SVDD过程可以去除离群点,提高字典质量。在小波域中将各频带分开重建,可避免低频图像中包含的不可靠高频信息对超分辨率结果的影响,从而恢复出可靠的高频信息。
[Abstract]:Aim in order to improve the robustness of image super-resolution algorithm to singular points of data, an image super-resolution reconstruction algorithm (Kmeans-SVDD) based on K-means clustering and support vector data description is proposed. Methods the training process: firstly, the approximate subbands of the training images are divided into several classes by K-means clustering algorithm, and then the singular points of each class are removed by using the support vector data description. Finally, the approximate subband and detail subband dictionary are trained in wavelet domain by principal component analysis (PCA). Testing process: according to the similarity of the approximate sub-bands in the same scene, the approximate sub-bands of the low-resolution test images to be reconstructed are regarded as the approximate sub-bands of the corresponding high-resolution test images. Then the detailed sub-bands of the high-resolution test image are recovered from the dictionary obtained by the training, and the high-resolution test image is obtained by inverse wavelet transform. Results compared with the current bicubic interpolation, Zeyde,ANR and Kmeans-PCA algorithms and Kmeans-SVDD algorithm improved the average PSNR of high resolution test images by 1.82 dB,0.37 dB,0.30 dB,0.15 dB. in turn. Conclusion A large number of experiments show that adding SVDD before dictionary training can remove outliers and improve dictionary quality. The reconstruction of each frequency band in wavelet domain can avoid the influence of the unreliable high-frequency information contained in the low-frequency image on the super-resolution result and thus restore the reliable high-frequency information.
【作者单位】：西安邮电大学通信与信息工程学院;
【基金】：国家自然科学基金项目(61340040)~~
【分类号】：TP391.41

【参考文献】