Learning Level Set Methods for Image Segmentation and Object Extraction


Arcot Sowmya.

Student: Xiongcai Cai.
Introduction: During the past decades, a wide variety of mathematical and computational frameworks have been proposed to deal with Computer Vision problems. In general, most of those problems can be formulated as data partitioning problems, particularly frame partitioning problems in Computer Vision such as image and video segmentation, visual object recognition and tracking.
Significance: Traditional methods in visual processing are based on spectral transforms or stochastic/statistical models. These methodologies have been highly successful. Recently, there has been increased interest in novel ways of analysing, formulating and representing the data partitioning problems via variational approaches. Notable examples of such techniques are total variation regularised image restoration methods and variational level set methods for image segmentation. More recently, many researchers in computer vision and applied mathematics started applying this new approach to other problems. One of them is to combine level set based active contour models with non-linear regression models for motion tracking. Others attempt to consider the potential for solving the general data partitioning problem using variational methods, which poses new challenges in extending and applying level set based active contour models to machine learning problems. This approach offers a systematic treatment of geometric features of visual information, such as shapes, contours and curvatures, as well as temporal features such as shape and pixel changes. In the level set framework, this requires those features to be weighted and combined in the energy function. Therefore, information fusion and parameter tuning become vital for the successful application of level set methods.
Applications: The developed automatic parameter tuning and information fusion method embedded in the level set method framework has been employed for providing original solutions to image segmentation and object extraction in computer vision. Several experimental results for each of the above tasks are supplied, demonstrating the effectiveness of the proposed solutions and indicating the potential of the dynamic tuning and fusion model.
Challenges: We have developed a learnable framework for automatic parameter tuning and information fusion of level set methods. This framework utilises a search and learn strategy to parameter tuning, based on the wrapper approach for feature selection. Firstly, a searching method is employed to generate a set of training instances which contains a feature vector describing the configuration of the current processing and an optimal parameter vector for that configuration. This is done by testing a set of parameter value candidates, evaluating their performance and selecting the one with best performance using techniques such as Sequential Search or Genetic Search. Then the set of instances are fed into a machine learning procedure to learn how to choose optimal parameter values for a new dataset. Learning techniques such as SVM-Regression and Neural Networks can be utilised for this task. The information fusion problem can be solved in the same tuning framework. Firstly, based on the nature of the level set method, a variety of features can be fused using a linear combination approach, where each feature is multiplied by a weighting parameter value and then combined to create the energy function of the level set method. By this approach, the information fusion problem becomes the parameter tuning problem: how to tune the weighted parameter to force the evolving contour to stop at the true boundaries of the objects. This can then be solved using the automatic parameter tuning framework described above.
top of page