AI Technique

基于Danbooru 2018同人画师数据集构建的97万张图、7万个二次元角色的识别数据集。使用动漫头像检测模型过滤单角色图片，为角色识别研究提供训练数据。

基于Yolo V3的动漫头像检测模型，仅需200张标注数据即可在多种番剧、线描、背面等场景下实现较高检出率。开源代码和模型权重，支持开箱即用。

Implementing Loopy Belief Propagation as a more general optimization approach for MRF, supporting non-grid graphs and non-binary cases, with image restoration experiment results.

Deep dive into what energy functions can be minimized via graph-cut, with the necessary and sufficient condition f(0,0)+f(1,1)≤f(0,1)+f(1,0) and application to image segmentation.

Introducing how to formulate image restoration as a Markov Random Field optimization problem and solve it using graph-cut algorithm to achieve global optimal results efficiently.