條件式建模與多層表示式在影像分析之應用

Abstract

條件機率建模技術近來被廣泛使用在高維度建模中，生成式模型在影像分析可分成兩個元件:影像解讀分佈與影像資料條件分佈。兩者皆可利用條件機率建模技術與比例差調技巧來建構，然而往往特徵表示式在比例差調公式中太過複雜，導致不易建構其機率分布，也因此導致執行實際問題時的計算量太過龐大。 在這兩年期的計畫中，主要目標是要提出一套一般性適用的標準流程，用來表現多層式模型的特徵分布，使其局部特徵分布可以被一致性的推導或計算，我們將會測試我們新發研發的相對座標建模技術，並且應用於人臉辨識與識別上。接下來我想要進階化這個技術，使其可用於多層式模型，並得到分布表示公式在局部的一致性與可算性。我們也會應用我們的建模方法在Sanskrit筆跡研究上，並且更進一步建立更細節或更抽象的元件關係表示法，另一方面從計算的角度上來說，我們需要設法得到簡化式表式法，進而使我們能夠用序列辨識法自動做到簡至細計算，更進一步我們也要研究``x-px”問題，使得我們可以設計一套精簡化程式加速它的計算。

Conditional modeling techniques have been widely used for modeling high dimensional data. In image analysis, generative models are typically made up of two parts, a prior distribution on image interpretations and a conditional distribution on image given its interpretation. Both of them can be established via conditional modeling and tilting techniques. However, the attribute representation in the tilting formula is often too complicate to design and to implement in real applications due to computation burden. The main goal of this two-year research plan is to propose a general procedure to obtain the representation of tilting attributes in hierarchical models such that the marginal distributions on local instantiations can be obtained or computed in a consistent way. We will first exam our recent discovery of novel relational coordinate modeling technique and apply it to face detection and recognition. Then we would like to generalize this technique to obtain a coherence representation in a hierarchy which can provide the same tilting formulation on its local instantiations (or at least computable formulation). We will apply our modeling technique to Sanskrit handwritten project and further to represent the detailed or abstract relationship of parts and objects. From computational points of view, we plan to build coarse representations so that we can automatically obtain a coarse-to-fine strategy via salience procedure. Moreover, in order to approximate the integration in detection formula, we will design a pruning algorithm which calls for further investigation on ``x-px” problem