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。

Objective To improve the robustness of the data singular points of super-resolution (SR) algorithm, we propose an SR algorithm based on K-means clustering and support vector data description (Kmeans-SVDD). Method The proposed algorithm is composed of two stages: training and testing. In the training stage, the approximating sub-band of trained images is clustered by the K-means method.SVDD is employed to drop singular points for each cluster. Finally, principle component analysis (PCA) is used to train the approximation and the detail sub-band dictionaries in the wavelet domain. In the testing stage, the approximating sub-band of the input low-resolution (LR) image is used as the approximating sub-band of the high-resolution (HR) image. This step is conducted on the basis of the observation that the approximating sub-band of the HR image is similar to it fo the corresponding LR image. Then, the prepared dictionaries are utilized to recover the detail sub-bands of the HR image. Finally, converse wavelet transformation is used to obtain the recovered HR image. Result The proposed Kmeans-SVDD algorithm is shown to be superior to existing algorithms with an average PSNR improvement of 1.82 dB, 0.37 dB, 0.30 dB, and 0.15 dB over the bicubic interpolation, Zeyde's algorithm, ANR algorithm, and K-means-PCA algorithm, respectively. Conclusion Extensive tests indicate that once the SVDD process is added before the dictionary training, the outliers can be removed and the quality of the dictionary can be finally improved. By reconstructing each band separately in the wavelet domain, the influences caused by the uncertain high-frequency information in low-frequency image to image super-resolution can be prevented; thus, reliable detail information can be restored.