|
|
|||||
|
|
|||||
|
ISAIAH™ (Image Scaling through Artificial Intelligence Acceleration
Hardware) In the particular field of image processing, recovering missing pixels
or scaling an image (generally towards a greater size or format) are
extensively performed operations. To determine the pixel value at a missing location, a classic method
consists of taking a small area of the data that is determined by a sampling
window and providing it to an interpolator. The interpolator uses a
polynomial equation with fixed degree and coefficients, which in turn
determine the filter coefficients. This standard method has some
inconveniences. Identification of adequate polynomial equations is not easy.
The coefficients are generally fixed once for all, and thus are clearly not
optimized for all the data or encountered if recalculated for each set of
data, consumes a large amount processing time. The number of coefficients is
limited (generally to less than 20), making the interpolation very
approximate in most cases. For example, if black and white type images are
processed, the interpolation may create grey levels that do not exist in the
original image. A great number of arithmetic operations have to be performed
that generally imply a number of serial computations, making the interpolation
very slow. Finally, the time for completing the interpolation depends on the
number of coefficients and is thus variable for each set of data. This application solves the above
problem by using an artificial neural network with Radial Basis Function
(RBF) architecture and Restricted Coulomb Energy (RCE) learning algorithm,
instead of arithmetic interpolators and extrapolators. In this NN model,
recognition and prediction tasks are performed using reference databases to
characterize input data. Depending upon the problem to be solved, these
reference databases contain patterns that are Sub-images, Sub-signals,
subsets of data, and combination thereof. The patterns that are stored in
these reference databases are called prototypes. They are represented by a vector,
i.e. an array in a space of a given dimension. Well-known methods for
characterizing new (unknown) patterns, referred to herein below as input
patterns, using reference databases are based upon input space mapping
algorithms such as the K-Nearest-Neighbor (KNN) or the Influence Field (IF)
based on vectors distance L-1. ISAIAH™ has been realized as a POC for
applications on images from satellites and for compression of images in 5G
trasmission. |
|
©2024_Luca_Marchese_All_Rights_Reserved Aerospace_&_Defence_Machine_Learning_Company VAT:_IT0267070992 NATO_CAGE_CODE:_AK845 Email:_luca.marchese@synaptics.org |
|
Contacts_and_Social_Media
|
