An exploration of robust model structures for image variations based on statistical approaches

本研究的目的是为图像识别找到对图像变化具有鲁棒性的通用模型结构。可分离晶格hmm （SL-HMM）是一种可以处理位置和尺寸变化的统计模型。本文提出了一种将基于概率主成分分析和因子分析的线性特征提取集成到sl - hmm模型中的新模型结构。此外，将sl - hmm扩展为能够表示旋转变化和包含显式状态持续时间模型的模型。

The objective of this research is to find general model structures which are robust to image variations for image recognition. Separable lattice HMMs (SL-HMM) have been proposed as statistical models which can deal with location and size variations. This research proposed a new model structure which integrates linear feature extraction based on probabilistic principal component analysis and factor analysis into SL-HMMs. Furthermore, SL-HMMs were extended to models which can represent rotation variations and which include explicit state duration models.