基于测量域二次采样的自适应压缩感知图像编码

Adaptive compressed sensing image code resampling in measurement domain

采用K-SVD训练字典代替DCT（discrete cosine transform）基,在测量域用K-means算法对原图像块根据显著性进行区域划分,对类别质心较高的测量值所对应的显著图像块分配较高采样率,对类别质心较低的测量值所对应的非显著图像块分配较低采样率,进行二次测量.实验结果表明,K-SVD字典可以较好地解决DCT基中存在的方块效应问题.并且,在相同稀疏基以及整幅图像同等压缩率的前提下,采用K-means二次测量算法可以显著提高图像的重构质量,PSNR值提高0.76-4.91dB.

DCT matrix is first replaced by the dictionary trained by K-SVD, and the image blocks are divided into salient regions and non-salient regions by K-means in measurement domain Then, the whole image is resampled by allocating the higher sam-pling rate to the salient regions, while lower sampling rate to the non-salient regions. Simulation results show that the dictionary trained by K-SVD solves the problem of poor profile representation Beside, with the same sparse matrix and compression ratio, K-means resampling method can improve the reconstruction quality that PSNR is improved by 0. 76-4. 91 dB.