特征驱动先验的归一化卷积超分辨率重建

目的 针对融合—复原法超分辨率重建中融合与复原两大环节，提出新的改进算法框架：用改进的归一化卷积实现融合，再用改进的最大后验估计实现复原，得到更优的超分辨率重建。方法 改进的归一化卷积引入了双适应度函数和一种新的混合确定度函数；改进的最大后验估计，引入一种特征驱动先验模型，该模型通过混合两种不变先验模型而得到，形式完全取决于图像自身的统计特征。结果 用本文算法对不同降质水平的图像进行重建，并与其他若干算法重建结果比较。无论从视觉效果还是从评价指标，本文算法均优于其他算法。结论 本文超分辨率重建算法，融合环节兼顾了邻域像素的空间距离和光度差，充分利用两种确定度函数的各自优势，可以抑制更多噪声和异常值；复原环节的先验模型依据图像特征而不是经验，对图像刻画更准确。实验结果也验证了本文算法的有效性。

Normalized convolution based super-resolution reconstruction using feature-driven prior

Objective Fusion restoration is one of the most concise and practical methods for resolution reconstruction. To solve the existing problems related to fusion and restoration, this study proposes a new improved framework. The normalized convolution is improved and then used to implement the fusion step. The maximum a posteriori estimation is improved and then used to implement the fusion step. These improvements lead to the construction of a super-resolution reconstruction algorithm. Method In the fusion step of the proposed algorithm, the improved normalized convolution introduces a double applicability function. That adds a neighbor-intensity correlation. Then the improved normalized convolution introduces a new certainty function that mixes the Gaussian and Laplace certainty functions. In the restoration step of the proposed algorithm, the improved maximum a posteriori estimation introduces a feature-driven function. This function is obtained by mixing two constant prior models. The formulation of the feature-driven prior is completely determined using the statistics of the image feature. Result Several test images are synthetically degraded into low-resolution sequences with different disturbance levels. These sequences are then reconstructed using the proposed algorithm and other several algorithms for comparison. Results show that the proposed algorithm is superior to other algorithms in terms of visual effects and performance indexes. Conclusion The fusion step in the proposed fusion restoration algorithm considers the spatial distance and intensity difference between neighboring pixels to efficiently restrain noise and outliers. The restoration step adopts the feature-driven prior that is determined by the image itself and not by experience. Therefore, the image is accurately characterized. The experimental results verify the effectiveness of the proposed algorithm.