为了以自动的方式达到反映原始场景真实性的图像复原效果,提出一个基于大位移视图的自动可信图像修补技术框架。首先,根据优缺点互补的原则将几种显著特征检测器有效地结合起来,提取目标图像和大位移视图上均匀分布的准稠密特征对应点集。然后,受启发于先验模型与模型拟合问题,提出一个准平面场景区域聚类算法,通过对特征对应点集的聚类划分将整个自然场景图像分割表示成多个准平面场景区域,以校正大位移视图中的景物投影变形。最后,受启发于纹理合成与图像拼接技术,提出一个准平面场景区域合成算法,校正并缝合空洞周围的多个准平面场景区域重投影图像至目标图像上,以填补目标图像上的信息丢失区域。实验结果与实拍照片之间的视觉辨别困难表明了本文方法的有效性。

In order to achieve accurate automatic image reparation, we propose a new framework for image completion based on the large displacement view(LDV)image. First, we extract the evenly distributed and quasi-dense feature-point correspondences between the target image and the LDV image by combining multiple distinct feature detectors in a complementary way. Inspired by the prior model and model fitting problem, we then devise a quasi-planar scene-regions(QPSRs)clustering algorithm, which classifies the feature point correspondences and represents a natural scene image with multiple QPSRs to remove the perspective distortions in the LDV image. Inspired by the texture synthesis and image stitching techniques, we finally present a QPSRs compositing algorithm, which corrects and stitches the re-projected QPSRs images to fill in the missing areas on the target image. Our experimental results are comparable with the ground-truth.