Improved Evolutionary Search for Image Reconstruction Transforms

Peterson, Michael R.; Lamont, Gary B.; Moore, Frances C.

doi:10.1109/cec.2006.1688671

Cited by 17 publications

(15 citation statements)

References 18 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…present an empirical comparison of various fitness functions and the application of local mutation operators to a GA for wavelet-based transform optimization. 5 Various mutation and crossover operators have been evaluated with respect to their suitability for image transform evolution. 12,13 Edge detection algorithms may be applied in conjunction with genetic algorithms to improve the resolution of structures within images that have been subjected to data loss due to quantization loss.…”

Section: Evolutionary Computation and Wavelet Optimizationmentioning

confidence: 99%

Image sets for satellite image processing systems

2011

View full text Add to dashboard Cite

The development of novel image processing algorithms requires a diverse and relevant set of training images to ensure the general applicability of such algorithms for their required tasks. Images must be appropriately chosen for the algorithm's intended applications. Image processing algorithms often employ the discrete wavelet transform (DWT) algorithm to provide efficient compression and near-perfect reconstruction of image data. Defense applications often require the transmission of images and video across noisy or low-bandwidth channels. Unfortunately, the DWT algorithm's performance deteriorates in the presence of noise. Evolutionary algorithms are often able to train image filters that outperform DWT filters in noisy environments. Here, we present and evaluate two image sets suitable for the training of such filters for satellite and unmanned aerial vehicle imagery applications. We demonstrate the use of the first image set as a training platform for evolutionary algorithms that optimize discrete wavelet transform (DWT)-based image transform filters for satellite image compression. We evaluate the suitability of each image as a training image during optimization. Each image is ranked according to its suitability as a training image and its difficulty as a test image. The second image set provides a test-bed for holdout validation of trained image filters. These images are used to independently verify that trained filters will provide strong performance on unseen satellite images. Collectively, these image sets are suitable for the development of image processing algorithms for satellite and reconnaissance imagery applications.

show abstract

Section: Evolutionary Computation and Wavelet Optimizationmentioning

confidence: 99%

Image sets for satellite image processing systems

2011

View full text Add to dashboard Cite

show abstract

“…In order to assess each image in terms of its usefulness as a training image and as a test image, we conduct one GA run for each image using standard operators that provided consistent solid performance in previous research [12]. Experiments are conducted using a GA previously demonstrating successful filter evolution [12,14]. The GA employs a population size of 50 evolved for 500 generations.…”

Section: Image Ranking Experimentsmentioning

confidence: 99%

“…An initial series of experiments identify the satellite images that when employed as GA training examples result in filters providing consistent reconstruction improvement over the entire set of satellite images. In order to assess each image in terms of its usefulness as a training image and as a test image, we conduct one GA run for each image using standard operators that provided consistent solid performance in previous research [12]. Experiments are conducted using a GA previously demonstrating successful filter evolution [12,14].…”

Section: Image Ranking Experimentsmentioning

confidence: 99%

“…By focusing on the evolution of optimized reconstruction coefficients, the underlying compression rate of the forward transform is unaffected. However, the resulting filters described in [11,9,12] are no longer wavelets because they no longer conform to the mathematical properties of wavelets, such as biorthogonality of the filters. Evolved with one or more training images, the resulting filters provide improved reconstruction when applied to images not explicitly represented by the training image population.…”

Section: Introductionmentioning

confidence: 99%

“…While GAs provide filters providing improved reconstruction over DWT filters in applications subject to quantization error [11,12], filters must be trained using images appropriate for the target application. An evolved transform filter trained using images of human faces may not provide sufficient reconstruction of an overhead image of a city captured from a high altitude.…”

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

A satellite image set for the evolution of image transforms for defense applications

Peterson

Lamont

Moore

et al. 2007

Proceedings of the 9th Annual Conference Companion on Genetic and Evolutionary Computation

View full text Add to dashboard Cite

In recent years, wavelets have been widely applied in stateof-the-art image processing algorithms, providing efficient compression while maintaining superior image quality. However, wavelet performance may not be sufficient when extreme compression ratios are required. Defense applications often require robust transforms simultaneously minimizing bandwidth requirements and image resolution loss. Image processing algorithms take advantage of quantization to provide substantial lossy compression ratios at the expense of resolution. Recent research demonstrates that genetic algorithms (GAs) evolve filters outperforming standard discrete wavelet transforms in conditions subject to high quantization error. Evolved filters must be trained using images appropriate to their intended application. We present a set of fifty satellite images used to evolve image transforms appropriate for satellite and unmanned aerial vehicle (UAV) reconnaissance applications. We identify the best training and test images. Image transforms evolved using appropriate training images reduce the mean squared error (MSE) by an average of greater than 15% across the entire image set under conditions subject to high quantization error.

show abstract