應用於影像分類的整體學習與轉移學習之新方法

New approaches of ensemble learning and transfer learning for image classificaion

王俊焜(Jun-Kun Wang)

指導教授 ： 鄭士康

國立臺灣大學/電機資訊學院/電信工程學研究所/碩士(2013年)

https://doi.org/10.6342/NTU.2013.00950

摘要

這篇碩論發展對照片分類有幫助的機器學習方法。我們針對整體學習與轉移學習這兩個領域各提出了新方法。在整體學習方面，我們使用了一包多向量支撐機辨識室內場景，此方法的概念簡單且實作容易。我們展示出有價值的視窗對室內場景辨識是重要的。根據我們的認知，這個影像特性過去是被忽略的。在轉移學習方面，我們給了一個嶄新的研究方向，找出所有任務都可方享的編碼機制。過去相關的研究都是關於怎麼利用任務的相關性來影響分類器的生成。我們的貢獻是提供一個新的研究方向，即利用任務的相關性來影響特徵的形成。我們提出督導階層式字典學習的架構，其可運作在轉移學習於多類別分類上。整個架構的運作是把經過編碼的特徵當作訊息在區塊間傳遞。分類器和字典透過這訊息來溝通。在我們的演算法中，向前和向後更新可視為在分類和編碼找出一個平衡點。

關鍵字

影像分類 ； 機器學習 ； 整體學習 ； 轉移學習 ； 字典學習

並列摘要

This thesis develops the machine learning approaches for image classification. Specifically, we consider two paradigms in machine learning, namely ensemble learning and transfer learning. In ensemble learning, we use a bag of
SVMs for indoor scene recognition, which is simple and easily implemented. We show valuable local windows are critical to scene recognition. To our knowledge, this image cue was ignored by the computer vision community previously. In transfer learning, we propose a new research direction that finds the encoding mechanism which all the tasks share. The common of all the related works is that they all deal with "how the task relatedness can affect the model during training". Our contribution is to provide another direction
that deal with "how the task relatedness can affect the features". The supervised hierarchical dictionary learning structure is proposed which works for multiclass classification with transfer learning. The whole architecture works by using encoding feature of training data as "messages" between different blocks. The models and the dictionaries are communicate by passing the messages.
The forward and backward updating in our algorithm can be view as trying to find a balance between classification and encoding.

並列關鍵字

image classification ； machine learning ； ensemble learning ； transfer learning ； dictionary learning

參考文獻

[AEP07] Andreas Argyriou, Theodoros Evgeniou, and Massimiliano Pontil. Multitask feature learning. In NIPS, 2007.

[BBLP10] Y-Lan Boureau, Francis Bach, Yann LeCun, and Jean Ponce. Learning midlevel features for recognition. In CVPR, 2010.

[Bis06] Christopher M. Bishop. Pattern recognition and machine learning. Springer,2006.

[Bre96] Leo Breiman. Bagging predictors. Machine Learning, 24(2), 1996.

[BRF12] Liefeng Bo, Xiaofeng Ren, and Dieter Fox. Unsupervised Feature Learning for RGB-D Based Object Recognition. In ISER, 2012.