新型高阶马尔科夫随机场模型在倾斜摄影图像的恢复算法

A Novel High-order Markov Random Field Model for the Recovery of Tilt-photography Images

提出一种基于深度学习网络结构的高阶马尔科夫随机场模型,充分考虑了彩色倾斜摄影图像色彩复杂的概率分布特征,设计并引入了概率特征可随实际像素变化的专家函数组,使该模型能更好地发现并获取影像图像像素间的自相似性.在深度学习网络的构建中,引入了非局域化优化算法,不仅使其能量函数的梯度形式大大简化,而且使该模型的优化过程的收敛效率得到了极大提高.与当今主流同类图像恢复算法进行对比,提出的模型不仅在平均倾斜摄影照片恢复效果上与当下最新同类算法在统计学上相当,甚至在某些色彩为主导的影像上性能要优于对方,而且极大简化了深度学习过程中的计算量和运算时间.

A high-order Markov random field model based on deep learning network structure is proposed,which fully considers the probability distribution characteristics of complex colors in color oblique shadow images.An expert function group with probability features that can vary with actual pixels is designed and introduced,which enables the model to better discover and obtain self-similarity between image pixels.In the construction of deep learning networks,non-local optimization algorithm is introduced,which not only greatly simplifies the gradient form of its energy function,but also greatly improves the convergence efficiency of the optimization process of the model.Compared with mainstream image restoration algorithms of the same kind today,the model proposed in this article not only has an average oblique photographic photo restoration effect,but also is statistically equivalent to the latest algorithms of the same kind.It even outperforms the other in some color-dominated images,and greatly simplifies the computational amount and operation time in the deep learning process.