楊佳玲臺灣大學：資訊網路與多媒體研究所洪銘蔚Hung, Ming-WeiMing-WeiHung2007-11-272018-07-052007-11-272018-07-052007http://ntur.lib.ntu.edu.tw//handle/246246/58418Retrieving images by textual queries requires some knowledge of the semantics of the image. Hence we need to find the label words that describe the content of the image and take them as the annotation of the image. Here, we propose an approach to annotate images with user feedback, and it annotates a label a time. The process contains a loop, and it will report a number of images which are most likely to be associated with the label word for user to annotate every iteration. The way to estimate the possibility that an image is associated with a label is using the known labeled images and some unlabeled images as training data to train a classifier for the label. While training the classifier, we use the semi-supervised learning method with unlabeled images to build hierarchical classifiers. The unlabeled images can help clustering while we only have a few labeled training images. After training the classifier, we take the unlabeled image as the input of the classifiers to estimate the confidence values representing the possibility that the image is associated with the label. After using the approach with every label words, we can get the annotation from all of the label words.1.Introduction 1 1.1Motivation 3 1.2 Overview 5 2. Related Work 9 2.1 Related Work of Automatic Image Annotation 9 2.2 Related Work of Semi-Supervised Learning 11 3. Background 13 3.1 Watershed Segmentation 13 3.2 k-means Clustering 15 3.3 Visual-word-based Image Representation 16 3.3.1 Visual Words 17 3.3.2 The Generation of Image Feature 18 3.4 Sift Descriptor 20 3.5 pLSA 21 4. Classifier Training 23 4.1 Stopping Condition 27 4.2 Score Function 28 4.3 Splitting Method 30 5. Confidence value 31 6. Experiments 33 6.1 Datasets 33 6.2 Experimental Results 34 7. Conclusion and Future Work 39 8. Reference 40790397 bytesapplication/pdfen-US影像標註影像檢索半指導式機器學習方法階層式分類器使用者反饋Image annotationImage retrievalsemi-supervised learninghierarchical classifieruser feedbackthesishttp://ntur.lib.ntu.edu.tw/bitstream/246246/58418/1/ntu-96-R94944019-1.pdf