Content Based Image Retrieval with Bag of Visual words

Content based image retrieval (CBIR) has become a popular area of research for both computer vision and multimedia communities. It aims at organizing digital picture archives by analyzing their visual contents. CBIR techniques make use of these visual contents to retrieve in response to any particular query. Note that this differs from traditional retrieval systems based on keywords to search images. Due to widespread variations in the images of standard image databases, achieving high precision and recall for retrieval remains a challenging task. In the recent past, many CBIR algorithms have applied Bag of Visual Words (BoVW) for modeling the visual contents of images. Though BoVW has emerged as a popular image content descriptor, it has some important limitations which can in turn adversely affect the retrieval performance.

Image retrieval has many applications in diverse ¿elds including healthcare, biometrics, digital libraries, historical research and many more (da Silva Torres and Falcao, 2006). In the retrieval system, two kinds of approaches are mainly followed, namely, Text-Based Image Retrieval (TBIR) and Content-Based Image Retrieval (CBIR). The former approach requires a lot of hu- man effort, and time and perception. Content based image retrieval is a technique that enables an user to extract similar images based on a query from a database containing large number of images.The basic issue in designing a CBIR system is to select the image features that best represent the image content in a database. As a part of a CBIR system, one has to apply appropriate visual content descriptors to represent these images. A query image should be represented similarly. Then, based on some measures of similarity, a set of images would be retrieved from the avail- able image database. The relevance feedback part, which incorporates inputs from a user, can be an optional block in a CBIR system. The fundamental problem in CBIR is how to transform the visual contents into distinctive features for dissimilar images, and into similar features for images that look alike.

BoVW has emerged as a popular model for representing the visual content of an image in the recent past. It tries to bridge the gap between low level visual features and high-level semantic features to some extent.