A Study of Image Retrieval and Relevance Feedback

• Main Authors: Yi-Lain Lin, 林宜聯
• Other Authors: Cheng-Chin Chiang
• Format: Others
• Language: zh-TW
• Published: 2004
• Online Access: http://ndltd.ncl.edu.tw/handle/13129725875136801854

碩士 === 國立東華大學 === 資訊工程學系 === 92 === Content-Based Image Retrieval (CBIR) takes visual features, instead of textual annotations, of images as the keys to retrieval images. This thesis presents a novel feature called “frequency layer” that reveals both the color and shape of image contents with respect to different frequencies of textures from the perspective of human vision. As every user may have his own subjective perception on the same image contents, each extracted key of visual feature should be of different importance for different users. This thesis also proposes two methods of relevance feedback to handle this problem of users’ subjectivity. One is designed on a neural network (NN) approach, while the other is based on the technique of maximum likelihood estimation (MLE). In to the experimental results in this research work, the performance of both methods are evaluated and compared. The results on testing some image databases show that both methods are effective. Moreover, the MLE method outperforms the NN approach.