IMPORTANT:

The documents contained herein are posted as a means to ensure timely dissemination of scholarly and technical work on a non-commercial basis. Copyright and all rights therein are maintained by the authors or by other copyright holders, notwithstanding that they have offered their works here electronically. It is understood that all persons copying this information will adhere to the terms and constraints invoked by each author's copyright. These works may not be reposted without the explicit permission of the copyright holder.

 

Papers, Patents, and Technical Reports:

 

Inventors: Brian Funt, Vlad Cardei and Kobus Barnard. Issued May 25, 1999.

A method of estimating the chromaticity of illumination of a colored image consisting of a plurality of color-encoded pixels.

 

by Vlad C. Cardei and Brian Funt

We show how to achieve better illumination estimates for color constancy by combining the results of several existing algorithms. We consider committee methods based on both linear and non–linear ways of combining the illumination estimates from the original set of color constancy algorithms. Committees of gray world, white patch and neural net methods are tested. The committee results are always more accurate than the estimates of any of the other algorithms taken in isolation.

 

by Vlad C. Cardei, Brian Funt and Kobus Barnard

Color images often must be color balanced to remove unwanted color casts. We extend previous work on using a neural network for illumination, or white-point, estimation from the case of calibrated images to that of uncalibrated images of unknown origin. The results show that the chromaticity of the ambient illumination can be estimated with an average CIE Lab error of 5DE. Comparisons are made to the gray world and white patch methods.

 

by Vlad C. Cardei, Brian Funt and Michael Brockington

Color correcting images of unknown origin (e.g. downloaded from the Internet) adds additional challenges to the already difficult problem of color correction, because neither the pre-processing the image was subjected to, nor the camera sensors or camera balance are known. In this paper, we propose a framework of dealing with some aspects of this type of image. In particular, we discuss the issue of color correction of images where an unknown ‘gamma’ non-linearity may be present. We show that the diagonal model, used for color correcting linear images, also works in the case of gamma corrected images. We also discuss the influence that unknown sensors and unknown camera balance has on color constancy algorithms.

 

by Vlad Cardei

This paper explores the possibility of recovering a lost color channel from an RGB image, based on the information present in the remaining two color channels, as well as on a priori knowledge about the statistics of sensor responses in a given environment. Different regression and neural network recovery methods are compared and the results show that even simple linear techniques suffice to obtain a good approximation of the original color channel.

Remote sensing and other visualization applications can benefit from these methods. Desaturated colors and skin tones are faithfully restored, which is important for the compression of still images and video signals.

Check out the related images that were not included in the final version of the paper:

      

 

by Brian Funt and Vlad Cardei

Bootstrapping provides a novel approach to training a neural network to estimate the chromaticity of the illuminant in a scene given image data alone. For initial training, the network requires feedback about the accuracy of the network’s current results. In the case of a network for color constancy, this feedback is the chromaticity of the incident scene illumination. In the past, perfect feedback has been used, but in the bootstrapping method feedback with a considerable degree of random error can be used to train the network instead. In particular, the gray world algorithm, which only provides modest color constancy performance, is used to train a neural network which in the end performs better than the gray world algorithm used to train it.
 
 

by Robert F. Hadley and Vlad C. Cardei

A connectionist-inspired, parallel processing network is presented which learns, on the basis of (relevantly) sparse input, to assign meaning interpretations to novel test sentences in both active and passive voice. Training and test sentences are generated from a simple recursive grammar, but once trained, the network successfully processes thousands of sentences containing deeply embedded clauses. All training is unsupervised with regard to error feedback -- only Hebbian and self-organizing forms of training are employed. In addition, the active--passive distinction is acquired without any supervised provision of cues or flags (in the output layer) that indicate whether the input sentence is in active or passive sentence. In more detail: (1) The model learns on the basis of a corpus of about 1000 sentences while the set of potential test sentences contains over 100 million sentences. (2) The model generalizes its capacity to interpret active and passive sentences to substantially deeper levels of clausal embedding. (3) After training, the model satisfies criteria for strong syntactic and strong semantic systematicity that humans also satisfy. (4) Symbolic message passing occurs within the model's output layer. This symbolic aspect reflects certain prior language acquisition assumptions.

 

by Vlad C. Cardei, Brian Funt and Kobus Barnard

In this article we present results that show that a multi-layer neural network can be trained to estimate the chromaticity of the scene illumination. The network is trained with a set of artificially generated scenes and the corresponding illuminants that were used for generating the scenes. In the test phase, the network estimates the illuminant under which a given scene was taken. Tests with artificially generated as well as natural scenes show that the neural network outperforms many current color-constancy algorithms and that it is more stable and reliable for scenes that contain a small number of colors. 

by Robert F. Hadley and Vlad C. Cardei
(check the abstract and the extended version of the paper - 110k Technical Report)

 

by Brian Funt, Vlad Cardei, Kobus Barnard

We previously developed a neural network which estimates the chromaticity of the illumination under which an given image was taken. This provides color constancy since, given the chromaticity estimate, the image pixel chromaticities can be converted via  a diagonal transformation to what they would be under a canonical illuminant. In tests on synthetically generated and real scene images, the accuracy of the illumination-chromaticity estimate generally surpassed that of most existing color constancy algorithms; however, the errors obtained with real images were significantly larger than those for the synthetic ones. After experiments with adding noise to the synthetic data, we concluded that there was a more fundamental problem than simply the influence of noise which remained to be explained. We hypothesized that specular reflection was causing the problem, so we modeled the specular reflection in the training set. The errors dropped by more than 20%.

 

by Brian Funt, Kobus Barnard, Michael Brockington, Vlad Cardei

Multi-scale retinex (MSR) processing has been shown to be an effective way to enhance image contrast, but it often has an undesirable desaturating effect on the image colors. A color-restoration method can help mitigate this effect, but our experience is that it simply leads to other problems. In this paper we modify MSR so that it preserves color fidelity while still enhancing contrast. We then add neural-net based color constancy processing to this modified version of MSR. The result is a principled approach that provides the contrast-enhancement benefits of MSR and improved color fidelity.]

by Vlad C. Cardei, Brian Funt and Kobus Barnard

 

by Robert F. Hadley and Vlad C. Cardei

A connectionist-inspired, parallel processing network is presented which learns, on the basis of (relevantly) sparse input, to assign meaning interpretations to novel test sentences in both active and passive voice. Training and test sentences are generated from a simple recursive grammar, but once trained, the network successfully processes thousands of sentences containing deeply embedded clauses. All training is unsupervised with regard to error feedback - only Hebbian and Kohonen forms of training are employed. In addition, the active-passive distinction is acquired without any supervised provision of cues or flags (in the output layer) that indicate whether the input sentence is in active or passive sentence.

 

by Brian Funt, Vlad Cardei and Kobus Barnard

We decided to test a surprisingly simple hypothesis; namely, that the relationship between an image of a scene and the chromaticity of scene illumination could be learned by a neural network. If a network could be trained to determine a scene's illumination only from the pixels in the image, it would then allow for the correction of the image colors to those relative to a standard illuminant, thereby providing color constancy. Using a database of surface reflectances and illuminants, along with the spectral sensitivity functions of our camera, we generated thousands of images of randomly selected illuminants lighting `scenes' of 1 to 60 randomly selected reflectances. During the learning phase the network is provided the image data along with the chromaticity of its illuminant. After training, the network outputs (very quickly) the chromaticity of the illumination given only the image data. We obtained surprisingly good estimates of the ambient illumination lighting from the network even when applied to scenes in our lab that were completely unrelated to the training data.
 
 

Research projects:

(All of them are in compressed Postscript format)

 

This project presents a hybrid neural network which can predict a set of microfeatures of the words from a given sentence. The idea motivating the whole project was to build a network that has a cognitive plausible structure the learning algorithms are unsupervised and the connections between neurons are only feed-forward and, using this network, to try to predict the microfeatures that correspond to the next word in a given sentence.

Similar work has been done by J. L. Elman who proved that connectionist representations can indeed possess internal structure and enable systematic behavior. The networks he designed predicted the lexical categories and dealt with relative clauses, but they had a rather complicated recurrent structure and supervised learning algorithms were used.

 

This project presents a brief history of fractals (focusing on fractal curves) and presents the fractal dimension as a means to characterize fractal shapes.