一种视差域基于学习方法的图像合成算法

基于图像序列的图像合成往往需要定标和恢复场景三维结构,为了避免这些复杂过程,根据参考图像的视点离虚拟视点的远近关系,将参考图像序列分为主、从参考图像,并利用多极约束的原理和非线性校正,将图像合成的最优化问题从深度域转化到视差域,从而在不定标和不匹配的情况下,直接从图像序列合成虚拟视图。实验结果证明算法是有效的,并有一定的应用前景。

View Synthesis Based on Learning in Disparity Field

The process of calibration and 3D reconstruction of the scene are always required in view synthesis from image sequences.To avoid those complicated processes,the reference images are first classified into primary reference image and subordinate reference image according to the distance between the reference viewpoints and the novel viewpoint.Then the global optimization problem of view synthesis is transformed from the depth field to disparity field,using a process of nonlinear rectification and multi-epipolar technology.Finally,the novel view is synthesized from the image sequence without matching and calibration.Experimental results show that the method is effective and has potential in the future.