Provided by Cognitive Sciences ePrint Archive
Accurate and robust image superresolution by neural
processing of local image representations
Carlos Miravet1,2 and Francisco B. Rodriguez1
1 Grupo de Neurocomputacidn Biologica (GNB), Escuela Politecnica Superior,
Universidad Autonoma de Madrid, 28049 Madrid, Spain
[email protected], [email protected]
2 SENER Ingenieria y Sistemas, S. A.,Severo Ochoa 4 (P.T.M.), 28760 Madrid, Spain
Abstract. Image superresolution involves the processing of an image sequence
to generate a still image with higher resolution. Classical approaches, such as
bayesian MAP methods, require iterative minimization procedures, with high
computational costs. Recently, the authors proposed a method to tackle this
problem, based on the use of a hybrid MLP-PNN architecture. In this paper, we
present a novel superresolution method, based on an evolution of this concept,
to incorporate the use of local image models. A neural processing stage receives
as input the value of model coefficients on local windows. The data dimension-
ality is firstly reduced by application of PCA. An MLP, trained on synthetic se-
quences with various amounts of noise, estimates the high-resolution image
data. The effect of varying the dimension of the network input space is exam-
ined, showing a complex, structured behavior. Quantitative results are presented
showing the accuracy and robustness of the proposed method.
Introduction
Image superresolution [1, 2] involves the processing of an image sequence to generate
a high-resolution description of the underlying scene. From the earliest algorithm pro-
posed by Tsai and Huang [3], a number of approaches have been proposed. Of these,
bayesian MAP (Maximum A Posteriori) methods [4,5] have gained particular
acceptance due to their robustness and their capability to incorporate a priori con-
straints. The main drawback of these methods comes from their associated high com-
putational loads, as they use iterative techniques in spaces of high dimensionality.
Recently [6, 7], the authors have proposed a neural network based technique that
provides results comparable to classical methods with a substantial decrease in com-
putational complexity. This technique estimates image values in a dense grid using an
irregular interpolation scheme, with distance dependent interpolation weights. Opti-
mal distance-to-weight mappings are learned from synthetic sequences and corre-
sponding high-resolution images, using a hybrid MLP-PNN (Multi Layer Perceptron
- Probabilistic Neural Network) architecture. In a second step, high-resolution image
values are restored from estimated grid values using optimal filters.
The use of interpolation schemes based on distance dependent weights, no matter
how optimally these weights can be tuned, poses a fundamental limit on the attainable