Real-time interactive image segmentation

We present a real-time interactive image segmentation system for the semantic analysis of images. It is a feasible solution to a variety of real quality control tasks mainly because it adopts a novel interactive training and classification mode. In general, our system incorporates the interactive acquisition of training data into the automatic classification based on adaptive boosting.

To reduce the manual effort required for interactively acquiring the online training data, the input images are efficiently processed through GPU-based over-segmentation. We also developed a user-friendly graphical user interface to allow users to edit (select, deselect and reselect) the training data. It also allows the users to add more training data to improve the classifier if necessary.

The experiments demonstrate that our system efficiently delivers a semantic analysis for images. More importantly, different from the state-of-the-art methods which typically rely on offline training using a large collection of data, our system can rely only on a tiny amount of training data offered by the users in an interactive manner and achieve a highly accurate classification.