结合多特征的单幅图像超分辨率重建算法

Single image super-resolution reconstruction based on multi-feature fusion

摘要:

为提高直接捕获的图像质量，针对梯度特征只能提取水平、垂直方向信息及非下采样轮廓波变换(NSCT)提取细节信息不足的缺陷，提出一种结合Gabor变换及NSCT的超分辨率重建算法.该算法充分利用Gabor变换和NSCT的互补性，针对输入图像块的特点，采用Gabor变换来提取纹理特征，NSCT来提取轮廓特征，然后分别利用稀疏模型进行重建，最后合并成一幅高分辨率图像.由于输入图像或多或少存在模糊，在重建过程中，加入了去模糊的正则项，以消除输入模糊的影响.实验结果表明，结合两种特征的超分辨率效果与单一特征相比，能够恢复更多的细节信息，去模糊正则项也有一定的作用.本文方法与Kim提出的核岭回归及Yang提出的稀疏表示算法(SCSR)相比，主观上视觉效果更加清晰，客观上PSNR值平均提高了近2dB，说明了该算法能够有效地提高图像的质量.

Abstract:

The gradients extract the information only along the horizontal and vertical directions and the non-subsampled contourlet transform (NSCT) is poor relatively to capture the detailed information. To overcome the drawback, a novel super-resolution approach combined Gabor with NSCT is proposed to improve the quality of image captured directly. The algorithm makes full use of the complementary of the Gabor transform and NSCT, to extract the texture feature using the Gabor transform and to extract the contour feature using the NSCT according to the characteristics of input image pieces. After that the sparse coding reconstruction is performed, and finally merge the pieces into a initial high-resolution image. Since the input image is blurred more or less, the approach revises the initial high-resolution image through the deblurred regularization to eliminate the influence of blurred input. Experiment results show that combining the Gabor and NSCT can recover more details and the deblurred regularization is also effective. Compared to the kernel ridge regression method proposed by Kim and the sparse coding super-resolution (SCSR) method proposed by Yang, the images produced by our approach are clearer in subjectively and the average PSNR is nearly 2 dB higher, which means that the proposed approach can improve the quality of image.