主题学习和稀疏表示的MODIS图像超分辨率重建

针对MODIS图像分辨率受传感器限制和噪声干扰,且分辨率局限在一定水平等问题,提出一种采用主题学习和稀疏表示的MODIS图像超分辨率重建方法,该方法通过双边滤波将MODIS图像的平滑及纹理部分分离,并将纹理部分看成是由若干"文档"组成的训练样本;运用概率潜在语义分析提取"文档"的潜在语义特征,从而确定"文档"所属的"主题"。在此基础上,针对每个主题所对应的图像块,采用改进的K-SVD方法训练若干适用于不同主题的高低分辨率字典对,从而可以运用这些字典对,通过稀疏编码实现测试图像相应主题块的超分辨率重建。实验结果表明,重建图像在视觉效果和PSNR等指标上均优于传统方法。

Super resolution reconstruction of MODIS image based on topic learning and sparse representation

MODIS images have important application value in the field of ground monitoring, cloud classification, and meteorological research. However, their image resolutions are still limited to a certain level because of the sensor limitations and external disturbance. This study attempts to reconstruct high-resolution MODIS images that make the edge clearer and more detailed by utilizing topic learning and the sparse representation method. The application value of existing MODIS images is then improved. A super resolution reconstruction method for MODIS images based on topic learning and sparse representation is proposed. The smoothing and texture parts of MODIS images are separated by the bilateral filtering method. The texture part is regarded as a training sample composed of several "documents". The latent semantic features of the "document" are extracted by probabilistic Latent Semantic Analysis (pLSA) to discover the inherent "topics" of "document". The improved K-SVD method trains several high- and low-resolution dictionary pairs that are suitable to different topics based on the aforementioned scenario, where the image blocks correspond to each topic. The probabilistic latent semantic analysis method is utilized in the reconstruction phase to adaptively select the image block topic, combine the dictionary of the corresponding topic, and reconstruct the high-resolution MODIS image through the sparse coding method. First, the MODIS image is blurred and subjected to down sampling processing in the experiment process to obtain a low-resolution image. Super resolution reconstruction is performed by utilizing different methods. The PSNR and SSIM of the original high-resolution and reconstructed images were compared utilizing different methods. Results show that the PSNR of the reconstructed image by our method is higher by approximately 1 dB and 0.5 dB than the bicubic interpolation and SCSR method, respectively. Its SSIM value is also higher than those of the other methods. The visual effects of super resolution reconstruction on the real images by different methods were compared. The experimental results show that the reconstructed images by our method have a high contrast ratio and rich texture details. The human vision is more sensitive to the image texture. This study separates the smoothing and texture parts of the MODIS image through the bilateral filter. The texture part is divided into multiple topics by probabilistic latent semantic analysis. A local adaptive super resolution method is constructed, which overcomes the problem of the adaptive selection of a reasonable dictionary according to the local characteristics of MODIS images. This process was conducted under the topic model framework combined with the improved K-SVD dictionary training methods, which train several high- and low-resolution dictionary pairs suitable to different topics. The experimental results show that the multi-dictionary reconstruction method can be utilized to represent MODIS images more sparsely and enhance the image reconstruction details. The experimental results also show that the reconstructed image is superior to the traditional method in terms of the visual effects, PSNR, and SSIM.