| Summary: | 碩士 === 朝陽科技大學 === 資訊管理系碩士班 === 95 === In recent years, digital photography has been gradually replacing traditional film photography into the mainstream imaging methods. Due to the convenience of digital photography, the average consumer may also possess a staggering amount of digital photographs. Therefore, how to effectively manage the consumer digital photo database has become a topic worthy of study. This paper is proposed a semantic-oriented consumer image retrieval system which combines low-level feature with high-level semantic to classify consumer images into five popular categories including humans, plants, transportations, buildings, and heavenly bodies. The system aims to extract the suitable color characteristic for each category and carry on the classification. In the image retrieval, the system provides user three ways to retrieve images. First is by the example image to carry on the retrieval. Second is to input the semantic vector to carry on the retrieval. Third is expected retrieval result can more near user intent through weight classification from the user’s interaction. We hope that the proposed image retrieval can avoid the original subjective consciousness of artificial retrieval and provide image retrieval of more adapted for perception of human. The experiments in the thesis use three color feature stages to perform semantic classification on 2000 genuine heterogeneous consumer photos. The precision of classification reaches 79.34%. There are three ways in the query process. The precision of query process reaches 84.37%.
|
|---|
