用階層圖像表示之一般物體辨識

Abstract

人和動物把生活中的事物分類之後，就可以用這類物體已有的知識做更好的決定和行動。我們希望讓機器也可以做到這件事。解決這個問題有兩個努力的方向，一是提出更好的圖像表示，另一個是用更佳的機器學習演算法，這篇論文選擇前者。 這篇論文先研究了數個之前用來表示一張圖像的方法。並且從中發現了數個建置圖像表示的共通運算原則。接著我們在這些法則之上。提出了一個新的前饋式階層圖像模型用來解決事物分類的問題。新模型的特點是有很大的參數可變性。可以根據要問題要求，用這個模型建出各種不同特性的圖像表示。 我們用這個模型建造了一個簡易四個階層的圖像分類系統。最後我們用 PASCAL VISUAL Object Challenge 2006的資料來測試這個模型的效果，最後我們得到了與一袋字模型(Bag of Word Models)相近的優良分類表現。

Humans and animals classify the general objects in daily l ife. After classifying the objects, they use the prior knowledge of this object class to do better decisions and actions. We hope one day the machines are also capable of this task. There are two directions in solving this problem. One is generating a more representative representation from images. Another is using better machine learning algorithms on the generated representation. This thesis walks toward the former direction. After surveying previous work, we discover a group of common operations in making a better representation. Based on these observations, a new feed -forward hierarchical model for image representation is proposed. This new model has the characteristi c of having large plasticity. The new model can fit the requirement of object classes by generating different descriptors. We build a simple 4-layers hierarchical image Categorization system to test this new model. This system is evaluated on the PASCAL Visual Object Challenge 2006 data set. And we get a similar performance of Bag of Words model.