摘要
目的解决当前方法需要对图像中的相应点手动标记界标,且局限于特定对象或形状变形的问题。方法提出一种可以同时实现图像颜色、外观和形态的图像低维表示算法。结果该算法通过将形态和外观的流形约束到低维子空间上,进一步降低了流形学习的采样复杂性。结论文中方法的性能远优于目前典型的稳健型光流算法和SIFT流算法。在图像编辑和关节学习关任务中取得了令人满意的定性结果。
The work aims to solve the problem that the existing solutions either require manually specified landmarks for corresponding points in the images, or are restricted to specific objects or shape deformations.A low-dimensional representation of images for simultaneously recovering color, appearance and shape was proposed.The proposed algorithm further reduced sample complexity of manifold learning as the manifolds of shape and appearance were restricted to low-dimensional subspaces.The proposed method significantly outperformed the current typical methods of robust optical flow and SIFT flow.Our qualitative results in some related tasks such as image deformation and joint learning are encouraging.
作者
曾步衢
ZENG Bu-qu(Huanghuai University, Zhumadian 463000, Chin)
出处
《包装工程》
CAS
北大核心
2017年第9期230-235,共6页
Packaging Engineering
基金
河南省教育厅重点科技攻关项目(13A520786)
