.png){kind=icon}
返回
Joint intensity-gradient guided generative modeling for colorization
.png){kind=icon}
PRE
AI
DOI：10.1007/s00371-022-02747-0.png)
.png){kind=icon}
.png){kind=icon}
分享.png){kind=icon}
收藏
摘要
En 中文
This paper proposes an iterative score-based generative model for solving the automatic colorization problem. Although unsupervised learning methods have shown the capability to generate plausible color, inadequate exploration of detailed information and data dimensions still limit the performance of the colorization model. Considering that the number of samples in score-based generative model has influence on estimating the target gradients and the gradient map possesses important latent information of the image, the inference process of the generative modeling is conducted in joint intensity-gradient domain for colorization. Specifically, a set of intensity-gradient formed high-dimensional tensors are trained, via the score matching, to attain the gradient of data distribution in joint intensity-gradient domain. As the score function is determined, data samples are generated by means of annealed Langevin dynamics, forming an iterative colorization procedure. Furthermore, the joint intensity-gradient constraint in data-fidelity term is proposed to limit the degree of freedom within generative model at the iterative colorization stage, thus being conducive to edge-preserving colorization effect. Experimental results conveyed the remarkable performance and diversity of our proposed method.
Keyword:
Automatic colorization
Unsupervised learning
Generative model
Intensity-gradient domain
Gradient constraint
期刊
IF:
2.9
论文数:
4.7K
被引数:
6.5K
机构
引用论文
.png){kind=icon}
Class diversification, economic growth and urban sprawl: evidences from a pre-crisis European city
PREAI
分享收藏
查看更多内容