提出一种基于TV-L1分解的红外云图超分辨率算法。该方法采用原始-对偶算法求解TV-L1图像分解模型, 将低分辨率云图分解为结构部分和纹理部分: 对结构部分采用软决策自适应插值(SAI)处理; 对纹理部分则基于非下采样Contourlet变换(NSCT)具有多方向和平移不变的特性, 构造非线性增益函数对其NSCT变换域系数进行处理, 然后对处理后的变换系数进行NSCT逆变换实现纹理增强。最后, 将处理后的结构部分和纹理部分组合起来得到重构的高分辨率云图。实验结果表明, 所提出的算法在视觉效果以及图像质量定量评价上均优于传统插值方法, 在实现两倍超分辨率时, 其峰值信噪比(PSNR)和结构相似度(SSIM)平均值分别提高了1.316 2~4.591 9 dB和0.007 1~0.020 6; 实现三倍超分辨率时PSNR和SSIM平均值分别提高了0.338 7~4.58 dB和0.001 8~0.041 7。由于SAI插值和非下釆样Contourlet变换准确表示了云图的不同形态特征, 故所提算法的超分辨率结果不但准确重建了云图中的结构部分, 而且有效保持了红外云图纹理和边缘。

Ａ super-resolution algorithm based on TV-L1 decomposition was proposed . In the algorithm, the original-dual algorithm was used to solve the TV-L1 image decomposition model, and the low resolution image was decomposed into the structure and the texture parts. The structure part was processed with a soft decision adaptive interpolation. For the texture part, the Nonsubsampled Contourlet Transform (NSCT) characterized by multi-direction and shift-invariance was used to construct the nonlinear gain function to process the NSCT transform domain coefficients, then the processed transform coefficients were enhanced their textures by the NSCT inverse transform. Finally, the reconstructed high resolution image was obtained by combining the processed the structure and texture parts. Experimental results show that the proposed algorithm in both visual effect and the quantitative evaluation on image quality is better than the traditional interpolation method. For realizing twice super-resolution, its peak signal-to-noise ratio (PSNR) and Structural Similarity (SSIM) average value are increased by 1.316 2-4.591 9 dB and 0.007 1-0.020 6. For realizing three super-resolution, the PSNR and the SSIM are increased by 0.338 7-4.58 0 dB and 0.001 8-0.041 7, respectively. Because of the accurate representation of the different morphological features of the cloud image, the SAI interpolation and NSCT not only reconstruct the smooth component, but also maintain the texture and edge of the infrared nephogram.