Abstract
Researchers at the University of Central Florida have developed an
innovative capsule-based deep learning system that requires far less parameterization to perform
object segmentation more accurately and efficiently than state-of-the-art convolutional
neural networks (CNNs) such as U-Net. The UCF deep
convolutional-deconvolutional capsule network called SegCaps also reduces the
cost, time and memory space needed to
automate the segmentation process.The system works with many computer
vision applications, such as medical imaging, for
effective cancer diagnosis.
In one
example application, UCF
researchers used the system's SegCaps network architecture to segment
pathological lungs from low-dose CT scans. Experimental results showed that compared to
the U-Net architecture, SegCaps reduced the number of parameters by 95.4 percent while still
providing better segmentation accuracy. Also,
the system aptly handled large image sizes (512 x 512 pixels) as well as baseline sizes (less than 32 x 32).
Technical Details
The invention comprises a
novel multi-task deep
convolutional-deconvolutional capsule architecture called SegCaps and methods
for using the system for improved, accurate object segmentation. To
dramatically reduce the memory and parameter burden of a capsule implementation,
SegCaps acts
locally when routing children
capsules to parent capsules and also allows for the segmentation of large image
sizes. It shares transformation matrices (sets of parameters for children
capsules) across capsules within the same capsule type. All of the capsule storage occurs
at the neuron level as vectors rather
than scalars. Vectors contain information about spatial orientation,
magnitude/prevalence, and other attributes of an extracted feature and
represent the "capsule types" within a layer.
The system's object segmentation process starts when a
computing device receives an input
image and passes it through a convolutional layer to produce feature maps. The maps
form a set of children
capsules with associated sets of parameters to model the spatial relationships
of objects. SegCaps uses the children capsules to create a set of prediction
vectors that are locally constrained within eachof the parent capsules'
kernels. The effort produces
locally accurate predictions for components of the input image, thus forming
meaningful part-to-whole relationships not found in standard CNNs.
Partnering Opportunity
The research team is looking for partners to develop the
technology further for commercialization.
Stage of Development
Prototype available.
Benefit
Reduces the space, time and costs associated with object segmentationAllows for a computing device to recognize images in a batch that are similar to an input imageProvides better accuracy than existing methods and requires 95 percent less parameterizationThe algorithm is usable in other platforms besides UNIXMarket Application
Medical imagingComputer visionImage segmentation for social mediaArtificial intelligenceMachine learningRobotics
Brochure