Today the amount of images that people have access to is large and steadily increasing which makes searching for an image a hard task. Traditional image browsing interfaces are based on timestamps or other kind of metadata such as GPS-tags or user defined captions.
This thesis presents a system which classifies images into the categories man-made/natural and indoor/outdoor which are found to be meaningful for users in an image browsing task. The system is implemented with state of the art feature extraction methods and classifiers, which are tested and evaluated. The distances between images in different feature spaces are used to create a combined distance matrix which is used in a final classification task. This classification achieves classifications rates of up to 93% for indoor/outdoor images.
To illustrate the potential use of such system the output of the classification system is used in an image browsing application. The image browsing application was developed for the iPad. In the application a user can filter and order a set of images based on their semantic content.
The system consisted of two parts. One for extracting features from images and one for classifying the image features.
The figure below illustrates the design of the classification system.
A scene classification system by it self is not very useful unless it can be incorporated into an application to enhance the user experience. Because of this an image browsing application was created for the iOS platform. In the application, the user could sort and filter images depending on the content of a semantic feature vector. A set of images were displayed in a grid that the user could interact with. Similarity search was done by tapping on an image, resulting in a reorder of the entire grid based on similarity to the selected image. Filtering of images was done by selecting a semantic class in a menu, resulting in that images that did not match the criterion's of that class were hidden.
The figure below is a screenshot from the image browsing application where the user has selected to only display images that have been classified as depicting nature scenes.
Semantic Scene Classification for Enhanced Image Browsing Experience,
Bladh, Gustav and Magnusson, Robert; Master Thesis, Lund University, April 2011.