基於非監督式特徵學習與監督式分類器學習演算法在手部切割之應用

Abstract

近年來特徵學習演算法已被廣泛地研究並且有許多應用，例如: 基準視覺識別問題: CIFAR-10以及STL資料庫。研究發現K-mean分群法可以被用作訓練特徵學習的快速替代方式，並能夠從影像處理中的未標記數據學習特徵。這種方法的主要優點是它在資料量大時能夠快速並且方便地實現。本研究的目標是以特徵學習演算法去擷取手的特徵，然後使用分類器學習完成手部切割。本文開發了基於非監督式特徵學習與監督式分類器學習演算法以完成手部切割的實現，其中，非監督式是使用K-mean分群法，而監督式是使用類神經網路。該過程包含兩個階段的學習:非監督式的特徵學習和監督式的分類器學習。第一階段包括三個步驟：前處理，產生patches和過濾器的學習。第二階段包括四個步驟：特徵提取，分類器學習，像素分類和後處理。我們的實驗顯示該演算法在手部分割的能力以及特徵學習演法能夠帶來更好的手部分割。

Recently, feature learning algorithm has been studied extensively for many applications, such as benchmark visual recognition problems: CIFAR-10 and STL datasets. It has been found that K-means clustering can be used as a fast alternative way to train features and is able to learn features from unlabeled data in images processing. The main advantage of this approach is that it is very fast and easily implemented at large scale. The goal of this research achieves feature learning algorithm to extract hand features, and then apply classifier learning to segment the hand from an image. This thesis develops unsupervised feature learning with K-means clustering and supervised classifier learning algorithms with neural network to achieve hand segmentation. The procedure contains two learning phases, unsupervised feature learning and supervised classifier learning phases. The first phase contains three steps: pre-processing, patches generation, and filters learning. The second phase contains four steps: feature extraction, classifier learning, pixels classification, and post-processing. Our experiments show the ability to segment hands with the proposed algorithm and verify that feature learning leads to better hand segmentation.