Abstract: Technologies for recommending relevant tags for the tagging of media based on one or more initial tags provided for the media and based on a large quantity of other tagged media. Sample media as candidates for recommendation are provided by a set of weak rankers based on corresponding relevance measures in semantic and visual domains. The various samples provided by the weak rankers are then ranked based on relative order to provide a list of recommended tags for the media. The weak rankers provide sample tags based on relevance measures including tag co-occurrence, tag content correlation, and image-conditioned tag correlation.

Abstract: Supervised re-ranking for visual search may include re-ordering images that are returned in response to a text-based image search by exploiting visual information included in the images. In one example, supervised re-ranking for visual search may include receiving a textual query, obtaining an initial ranking result including a plurality of images corresponding to the textual query, and representing the textual query by a visual context of the plurality of images. A query-independent re-ranking model may be trained based on visual re-ranking features of the plurality of images of the textual query in accordance with a supervised training algorithm.

Abstract: Techniques for construction of a visual codebook are described herein. Feature points may be extracted from large numbers of images. In one example, images providing N feature points may be used to construct a codebook of K words. The centers of each of K clusters of feature points may be initialized. In a looping or iterative manner, an assignment step assigns each feature point to a cluster and an update step locates a center of each cluster. The feature points may be assigned to a cluster based on a lesser of a distance to a center of a previously assigned cluster and a distance to a center derived by operation of an approximate nearest neighbor algorithm having aspects of randomization. The loop terminates when the feature points have sufficiently converged to their respective clusters. Centers of the clusters represent visual words, which may be used to construct the visual codebook.

Abstract: Technologies for recommending relevant tags for the tagging of media based on one or more initial tags provided for the media and based on a large quantity of other tagged media. Sample media as candidates for recommendation are provided by a set of weak rankers based on corresponding relevance measures in semantic and visual domains. The various samples provided by the weak rankers are then ranked based on relative order to provide a list of recommended tags for the media. The weak rankers provide sample tags based on relevance measures including tag co-occurrence, tag content correlation, and image-conditioned tag correlation.

Abstract: Techniques described herein create an accurate active-learning model that takes into account a sample selection bias of elements, such as images, selected for labeling by a user. These techniques select a first set of elements for labeling. Once a user labels these elements, the techniques calculate a sample selection bias of the selected elements and train a model that takes into account the sample selection bias. The techniques then select a second set of elements based, in part, on a sample selection bias of the elements. Again, once a user labels the second set of elements the techniques train the model while taking into account the calculated sample selection bias. Once the trained model satisfies a predefined stop condition, the techniques use the trained model to predict labels for the remaining unlabeled elements.

Abstract: A general framework for video search reranking is disclosed which explicitly formulates reranking into a global optimization problem from the Bayesian perspective. Under this framework, with two novel pair-wise ranking distances, two effective video search reranking methods, hinge reranking and preference strength reranking, are disclosed. Experiments conducted on the TRECVID dataset have demonstrated that the disclosed methods outperform several existing reranking approaches.

Abstract: A hierarchical sparse codebook allows efficient search and comparison of images in image retrieval. The hierarchical sparse codebook includes multiple levels and allows a gradual determination/classification of an image feature of an image into one or more groups or nodes by traversing the image feature through one or more paths to the one or more groups or nodes of the codebook. The image feature is compared with a subset of nodes at each level of the codebook, thereby reducing processing time.

Abstract: An ImageWiki architecture is used to generate an image-based web page for an image on the Web. An ImageWiki page may be created automatically or individually, by a user of the Web. Additionally, a user may revise existing ImageWiki pages to update a particular page or correct an incorrect or misleading previous entry. The ImageWiki application indexes images located on the Web. Once the images are indexed, the information related to the images is mined and extracted from various sources of web data. Finally, an ImageWiki page or web page is generated for each image. The resulting ImageWiki page contains the image as well as the aggregated information relating to the image.

Abstract: Technologies for generating a boosted tag ranking for a media instance, the boosted tag ranking based on probabilistic relevance estimation computed by a probabilistic relevance estimator and tag correlation refining performed by a tag correlation refiner. Such boosted tag rankings may be used for search result ranking, tag recommendation, and group recommendation.

Abstract: Adaptive image retrieval image allows retrieval of images that are more likely to reflect a current trend of user preferences and/or interests, and therefore can provide relevant results to an image search. Adaptive image retrieval includes receiving image query log data from one or more clients, and updating a codebook of features based on the received query log data. The image query log data includes images that have been queried by the one or more clients within a predetermined period of time.

Abstract: An adaptation process is described to adapt a ranking model constructed for a broad-based search engine for use with a domain-specific ranking model. An example process identifies a ranking model for use with a broad-based search engine and modifies that ranking model for use with a new (or “target”) domain containing information pertaining to a specific topic.

Abstract: Colors of images and videos are modified to make differences in the colors more perceptible to colorblind users. An exemplary recoloring process utilizes a color space transformation, a local color rotation and a global color rotation to transform colors of visual objects from colors which may not be distinguishable by the colorblind user to colors which may be distinguishable by the colorblind user.

Abstract: An informative priors image search result summarization system and method that summarizes image search results based on the image relevance (as determined by a search engine's initial ranking) and the image quality. Embodiments of the system and method cluster the image search results, rank images within each cluster based on a computed image score, and then select a summary image for the cluster. Each cluster is analyzed and an image in the cluster having the maximum image score is included in a selected summary collection. The image score is computed using the image relevance and the image quality, as well as a cluster coherence, a density, and a diversity. The selection of images from a collection of candidate images generates an image search result summarization, which is presented to a user. The summaries are presented to the user in a ranked order based on their image scores.

Abstract: This document describes techniques that utilize a learning method to generate a ranking model for use in image search systems. The techniques leverage textual information and visual information simultaneously when generating the ranking model. The tools are further configured to apply the ranking model responsive to receiving an image search query.

Abstract: Technologies for generating a boosted tag ranking for a media instance, the boosted tag ranking based on probabilistic relevance estimation and tag correlation refining. Such boosted tag rankings may be used for search result ranking, tag recommendation, and group recommendation.

Abstract: Technologies for recommending relevant tags for the tagging of media based on one or more initial tags provided for the media and based on a large quantity of other tagged media. Sample media as candidates for recommendation are provided by a set of weak rankers based on corresponding relevance measures in semantic and visual domains. The various samples provided by the weak rankers are then ranked based on relative order to provide a list of recommended tags for the media. The weak rankers provide sample tags based on relevance measures including tag co-occurrence, tag content correlation, and image-conditioned tag correlation.

Abstract: Techniques described herein create an accurate active-learning model that takes into account a sample selection bias of elements, such as images, selected for labeling by a user. These techniques select a first set of elements for labeling. Once a user labels these elements, the techniques calculate a sample selection bias of the selected elements and train a model that takes into account the sample selection bias. The techniques then select a second set of elements based, in part, on a sample selection bias of the elements. Again, once a user labels the second set of elements the techniques train the model while taking into account the calculated sample selection bias. Once the trained model satisfies a predefined stop condition, the techniques use the trained model to predict labels for the remaining unlabeled elements.

Abstract: Techniques described herein enable better understanding of the intent of a user that submits a particular search query. These techniques receive a search request for images associated with a particular query. In response, the techniques determine images that are associated with the query, as well as other keywords that are associated with these images. The techniques then cluster, for each set of images associated with one of these keywords, the set of images into multiple groups. The techniques then rank the images and determine a representative image of each cluster. Finally, the tools suggest, to the user that submitted the query, to refine the search based on user selection of a keyword and a representative image. Thus, the techniques better understand the user's intent by allowing the user to refine the search based on another keyword and based on an image on which the user wishes to focus the search.

Abstract: Colorblind accessible image search technique embodiments are presented that re-rank the results of a relevance-ranked image search to account for the accessibility of the images to a colorblind person. This is accomplished by first computing a colorblind accessibility quantity for each image of interest in the search results. A colorblind accessibility quantity quantizes the degree to which color information is preserved when an image is perceived by a colorblind person viewing the image. It is computed by generating a colorblind version of an image that simulates how the image would appear to the colorblind person. An amount quantifying the loss of color information between the image and the colorblind version of the image is then estimated. This estimate is used to compute the colorblind accessibility quantity for the image. Once the colorblind accessibility quantities have been computed, the image search results are re-ranked based on these quantities.

Abstract: Computer program products, devices, and methods for generating in-text embedded advertising are described. Embedded advertising is “hidden” or embedded into a message by matching an advertisement to the message and identifying a place in the message to insert the advertisement. For textual messages, statistical analysis of individual sentences is performed to determine where it would be most natural to insert an advertisement. Statistical rules of grammar derived from a language model may be used choose a natural and grammatical place in the sentence for inserting the advertisement. Insertion of the advertisement creates a modified sentence without degrading a meaning of the original sentence, yet also includes the advertisement as a part of a new sentence.