基于结构相似度的遥感图像质量评价

发布时间：2018-05-23 18:51

本文选题：遥感图像 + 全参考　；参考：《西安科技大学》2017年硕士论文

【摘要】：目前已经涌现了大量的图像质量评价算法且在评价普通图像时能够取得较好的效果。遥感图像作为一种高空拍摄的高分辨率自然图像,与普通的自然图像存在很多区别,采用现有的图像质量评价方法不能达到要求。因此就需要结合遥感图像本身的特点研究出更符合人眼视觉的质量评价方法。本文深入分析了结构相似性理论(Structural Similarity,SSIM),针对遥感图像边缘多尺度、纹理错综复杂的特点,结合人眼视觉系统,从全参考和无参考两个方面对SSIM方法进行改进。人眼对边缘区域及其两侧的对比度更加敏感,纹理区域本身包含整幅图像的大部分结构信息,然而SSIM将所有像素一视同仁,不符合人眼视觉系统。因此,本文给出了一种了全参考的基于边缘纹理区域质量评价算法(ET_SSIM)。该方法采用非下采样轮廓波变换进行多尺度分解,将子图像进行k-means边缘检测及细化融合,最终得到边缘区域。对图像计算其梯度幅值,设置两个阈值,将满足区间的像素提取出来得到纹理区域。然后,使用边缘计算对比度分量,纹理计算结构相似度分量,对SSIM进行改进。由于遥感图像缺少主观评价数据库,为了更好的检验本文方法的准确性,建立了遥感图像数据库进行验证。实验表明,本文的全参考方法ET_SSIM比MSE、PSNR、SSIM及GSSIM的线性相关度分别提高了 22.8%,6.7%,1.3%,0.7%。为了得到一种更加通用的无参考遥感图像评价方法,本文对ET_SSIM进行改进。采用“再降质”的方式构造出参考图像,将边缘纹理区域与平滑区域分开评价,ET_SSIM对边缘纹理区域评价,SSIM对平滑区域评价,根据人眼对边缘纹理及平滑区域的不同重视程度加权求和,得到最终结果。最后,在相同SSIM值不同主观感受的遥感图像集、自建数据库和LIVE数据库的实验说明本文方法VSSIM评价的单调性和一致性略低于ET_SSIM,但是作为一种无参考方法,其准确性明显好于SSIM。且本文的无参考方法VSSIM方法比目前主流的无参考方法BLIINDS-II线性相关度提高了 1.3%。
[Abstract]:At present, a large number of image quality evaluation algorithms have emerged and can achieve good results in the evaluation of ordinary images. As a kind of high-resolution natural image, remote sensing image is different from ordinary natural image, and the existing image quality evaluation method can not meet the requirements. Therefore, it is necessary to combine the characteristics of remote sensing images to develop a quality evaluation method that is more suitable for human vision. In this paper, the structural similarity theory is deeply analyzed. Aiming at the multi-scale and intricate texture of remote sensing image, the SSIM method is improved from two aspects: full reference and no reference. The human eye is more sensitive to the contrast of the edge region and its two sides, and the texture region itself contains most of the structural information of the whole image. However, SSIM treats all pixels equally and does not conform to the human visual system. Therefore, an all-reference quality evaluation algorithm based on edge texture region is presented. In this method, the non-downsampling contour wave transform is used for multi-scale decomposition, and the k-means edge detection and thinning fusion are performed on the sub-image, and the edge region is finally obtained. The gradient amplitude of the image is calculated and two thresholds are set to extract the pixels satisfying the range to obtain the texture region. Then, the contrast component is computed by edge and the structural similarity component is calculated by texture, and the SSIM is improved. Because remote sensing image lacks subjective evaluation database, in order to verify the accuracy of this method, a remote sensing image database is established for verification. The experimental results show that the linear correlation of the full reference method ET_SSIM is 22.86.7and 1.30.70% higher than that of MSE PSNRIM and GSSIM, respectively. In order to obtain a more general evaluation method of remote sensing images without reference, this paper improves ET_SSIM. The reference image is constructed by "redegrade" method, and the edge texture area is evaluated separately from the smooth area by ETS IM and the smooth area is evaluated by SSIM. The final results are obtained by weighted summation according to the different emphasis on edge texture and smooth region. Finally, in the remote sensing image set with the same SSIM value and different subjective feelings, the experiments of self-built database and LIVE database show that the monotonicity and consistency of VSSIM evaluation in this paper is slightly lower than that of ETSSS, but as a non-reference method, its accuracy is obviously better than that of SSIM. The linear correlation of the non-reference method VSSIM method is 1.3% higher than that of the current mainstream non-reference method BLIINDS-II method.
【学位授予单位】：西安科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP751

【参考文献】