Papers by Author: Zheng Liu

Abstract: In this paper, we concentrate on how to automatic detect landmarks of a city leveraging the community-contributed collections of rich media on the Web, as landmark for a given city could provide helpful information for tourist guides. Our approach only need the user to provide the city name, and then submit it to Flickr website to obtain photos and related metadata. Next, these Flickr photos are clustered by simultaneously integrating multiple types of metadata which are related to Flickr photos. Finally, landmarks are mined from the photos clustering results. Experiments conducted on the photos in Flickr demonstrate the effectiveness of the proposed approach and our approach could enhance the performance of tourist guiding systems greatly.

Abstract: We present an approach to tag image automatically via visual topic detecting and initial annotations expanding. Visual topics are detected from corel5k dataset by probabilistic latent semantic analysis (PLSA) model. For an image which is to be tagged, PLSA is used to find visual topic of this image, and then construct initial annotations set. After initial annotations are generated, we use a weighted voting scheme and Flickr API to expand initial annotations. After the above two process, we combine initial annotations and expanded annotations together to construct final annotations. From experimental results, the conclusions can be draw that our PLSA based image tagging approach works effectively.

Abstract: Traditional image clustering methods mainly depends on visual features only. Due to the well-known “semantic gap”, visual features can hardly describe the semantics of the images independently. In the case of Web images, apart from visual features, there are rich metadata which could enhance the performance of image clustering, such as time information, GPS coordinate and initial annotations. This paper proposes an efficient Flickr photo clustering algorithm by simultaneous integration information of multiple types which are related to Flickr photos using k-partite graph partitioning. For a personal collection of Flickr, we firstly determine the value of k which means the number of data types we used. Secondly, these heterogeneous metadata are mapped to vertices of a k-partite graph, and relationship between the heterogeneous metadata is represented as edge weight. Finally, Flickr photos could be clustered by partitioning the k-partite graph. Experiments conducted on the photos in Flickr demonstrate the effectiveness of the proposed algorithm.

Abstract: This paper presents LDA-based automatic image annotation by visual topic learning and related annotation extending. We introduce the Latent Dirichlet Allocation (LDA) model in visual application domain. Firstly, the visual topic which is most relevant to the unlabeled image is obtained. According to this visual topic, the annotations with highest likelihood serve as seed annotations. Next, seed annotations are extended by analyzing the relationship between seed annotations and related Flickr tags. Finally, we combine seed annotations and extended annotations to construct final annotation set. Experiments conducted on corel5k dataset demonstrate the effectiveness of the proposed model.

Abstract: Existing image annotation approaches mainly concentrate on achieving annotation results. Annotation order has not been taken into account carefully. As orderly annotation list could enhance the performance of image retrieval system, it is of great importance to rank annotations. This paper presents an algorithm to rank Web image annotating results. For an annotated Web image, we firstly partition the image by a region growing method. Secondly, relevance degree between two annotations is estimated through considering both semantic similarity and image content. Next, the regions of unlabeled image to be ranked serve as queries and annotations are used as the data points to be ranked. And then, manifold-ranking algorithm is executed to get the ordered annotation list. Experiments conducted on real-world Web images through NDCG metric demonstrate the effectiveness of the proposed approach.

Abstract: This paper presents a novel transport layer protocol for multi-level wireless sensor networks. The node of lower level uses a lightweight transmission protocol, which predigests the head of segment and a six-state finite state machine is applied. In order to make the highest nodes convenient for connecting with exterior networks, we modify TCP in the aspects of segment caching and local segment retransmissions, and use the TCP modified in wireless sensor networks directly. Simulation results show that our design can improve the communication capability of the transmission layer in multilevel wireless sensor networks greatly.