由于一些传统的超分辨率重建算法学习多幅不同类别的图像仍无法获得好的重建效果, 本文提出了一种基于支持向量回归机和光栅扫描的单帧图像超分辨率重建算法。该算法首先采用光栅扫描法对一组高低分辨率训练图像提取图像块, 从块中分别抽取输入向量和标签像素。利用Log算子判断这些块是属于高频空间还是低频空间, 从而构建高低频空间向量对并对其进行优化。然后, 用支持向量回归机(SVR)工具训练优化后的向量对, 得到高低频空间下的两个字典;抽取测试低分辨率图像中的块并得到高低频空间下的输入向量, 利用SVR工具回归对应的属于超分辨率图像块的标签像素并得到回归后的图像。最后, 对图像进行后处理得到最终的超分辨率图像。与其它算法的对比实验表明: 提出的算法具有较好的视觉效果。特别在放大倍数为2时, 提出的算法在不同图像上的峰值信噪比(PSNR)和结构相似度(SSIM)值较双三次插值法分别提高了3.1%~5.3%和1.5%~8.1%。得到的结果显示提出的算法获得了更好的重建效果。

Some of the traditional single-frame super-resolution (SR) reconstruction algorithms can not get good reconstruction results, although they learns many different types of images. Therefore, a super-resolution method combined with the Support Vector Regression (SVR) and raster-scan actions was proposed. Firstly, image patches were extracted from a group of high resolution (HR) images and the corresponding low resolution (LR) edition by the raster-scan actions, and input vectors and pixel vectors were taken out from the patches. The Log algorithm was used to determine that those patches were belong to high-frequency space or low frequency space then to construct the high and low frequency vector pairs. Then, those optimized vector pairs were trained by the SVR and two dictionaries in high/low frequency spaces were built eventually. Furthermore, input vectors were extracted from tested LR images in high/low frequency space, and the SVR tool was used to predict the SR pixel labels and the predicted pixels were added to bicubic interpolation image based on LR edition. Finally, the SR image was obtained by post-processing the previous image. In comparison with other algorithms, experimental results indicate that the proposed method provides good visual effects. It enhances its Peak Signal-to-Noise Ration (PSNR) and Structural Similarity Index Measurement (SSIM) by 3.1%-5.3% and 1.5%-8.1% on different images, respectively as compared with bicubic interpolation method.