SUMMARY
Researchers in the UCLA Department of Computer Science have developed an end-to-end trainable image segmentation framework that unifies Active Contour Models (ACM) with Convolutional Neural Network (CNNs) with learnable parameters.
BACKGROUND
The ACM (Active Contour Model) is a popular framework used to identify an object from a noisy 2D image. It has been successfully employed in various image analysis techniques including object segmentation and tracking. This framework is a model-based formulation founded on geometric and physical principles, which relies on the content of the image, large annotated image datasets, extensive computational resources, and hours or days of training to identify an object. This limits its application to automated analysis of large datasets since they rely on some degree of user interaction to specify the initial contour and tune the parameters. New ACM algorithms are needed to expand the application of ACM to automated image segmentation.
INNOVATION
Researchers in UCLA Department of Computer Science have introduced a novel image segmentation framework, called DCAC, which is a truly end-to-end integration of ACMs (Active Contour Model) and CNNs (Convolutional Neutral Network). The framework allows for pixel-wise learnable parameters that can adjust the contour to precisely capture and delineate the boundaries of objects of interest in the image. DCAC requires minimal human supervision and is initialized and guided by its CNN backbone. Moreover, DCAC can segment multiple objects simultaneously and is effective in handling various topological changes in the image.
POTENTIAL APPLICATIONS
ADVANTAGES
RELATED MATERIALS
DEVELOPMENT-TO-DATE
The framework has been successfully developed and used to identify building instance segmentation in aerial images.