Content Based Image Retrieval


  • Skills:

    • Java
    • Computer Vision
    • Machine Learning

BitTree, in collaboration with the group PRA Department. of Electrical and Electronic Engineering the University of Cagliari, is engaged in research in the field of databases based visual content querying (Content Based Image Retrieval CBIR).

The CBIR aims to recover one or more images from visual databases according to similarity with an image (query) the user chooses. This similarity is evaluated based on raw characteristics (such as color, shape, texture) extracted automatically from images. Many techniques and algorithms have been proposed in the literature to performing queries based on visual content in databases. Usually consist of presenting the user a set of images similar to image user query and using the user evaluation to refine research results . We are currently working on integrating an engine for CBIR query mechanisms with a digital library. Furthermore we are dealing with the development of advanced techniques for refining Querying of visual databases by integrating a semantic search.

BitTree dispose to a prototype to show the capabilities of a Content Based Image Retrieval system that based on the use of the Query by Example method together with a Relevance Feedback system. The proposed system is queried by a sample image (query by example) instead of using keywords or tags. Starting from this query image the system output the images more "similar" to the query image. The similarity degree is computed on the basis of visual contents of the images described by different features (e.g. texture, colors distributions, etc). In general, this fist stage doesn't provide high relevant results because it is an hard task to model the similarity semantic through low level visual features. At this stage the user can submit to the system a feedback, this feedback rappresentes the relevance of the image with respect of the semantic desidered. The system implements three degrees of relevance: relevant, not relevant, and neutral. These relevance degrees are subdued to the system that, after a computation, rearrange the researce previously made in a way to present to the user more relevant images. In this way it is possible to find similar images in tagged images databases, and, above all, in not-tagged images databases or with meaningless keyword.