Statistical models for text query-based image retrieval

Image indexing and retrieval has been an active research area for more than one decade. Although many accomplishments have been made in this domain, it is still a challenging problem and far from being solved. Traditional content-based approaches make use of queries based on image examples or image attributes like color and texture, and images are retrieved according to the similarity of each target image with the query image. However, image query based retrieval systems do not really capture the semantics or meanings of images well. Furthermore, image queries are difficult and inconvenient to form for most users. To capture the semantics of images, libraries and other organizations have manually annotated each image with keywords and captions, and then search on those annotations using text retrieval engines. The disadvantage of this approach is the huge cost of annotating large number of images and the inconsistency of annotations by different people. In this work, we focus on general image and historical handwritten document retrieval based on textual queries. We explore statistical model based techniques that allow us to retrieve general images and historical handwritten document images with text queries. These techniques are (i) image retrieval based on automatic annotation, (ii) direct retrieval based on computing the posterior of an image given a text query, and (iii) handwritten document image recognition. We compare the performance of these approaches on several general image and historical handwritten document collections. The main contributions of this work include (i) two probabilistic generative models for annotation-based retrieval, (ii) a direct retrieval model for general images, and (iii) a thorough investigation of machine learning models for handwritten document recognition. Our experimental results and retrieval systems show that our proposed approaches may be applied to practical textual query based retrieval systems on large image data sets.