R.D. Dony, S. Haykin, Neural network approaches to image compression, Proc. IEEE 83 (2) (February 1995) 288}303.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....possibilities of future research directions. 2. Direct neural network development for image compression 2.1. Back propagation image compression 2.1.1. Basic back propagation neural network Back propagation is one of the neural networks which are directly applied to image compression coding [9,17,47,48,57]. The neural network structure can be illustrated as in Fig. 1. Three layers, one input layer, one output layer and one hidden layer, are designed. The input layer and output layer are fully connected to the hidden layer. Compression is achieved by designing the value of K, the number of neurones ....

.... the hidden neurone output is real valued, quantization is required for xed length entropy coding which is normally designed as 32 level uniform quantization corresponding to 5 bit entropy coding [9,14] This neural network development, in fact, is in the direction of K L transform technology [17,21,50] which actually provides the optimum solution for all linear narrow channel type of image compression neural networks [17] When Eqs. 2.1) and (2.2) are represented in matrix form, we have [h] #[x] 2.4) xN ] #] h] #] #[x] 2.5) for encoding and decoding. The K L transform maps input ....

[Article contains additional citation context not shown here]

R.D. Dony, S. Haykin, Neural network approaches to image compression, Proc. IEEE 83 (2) (February 1995) 288}303.

.... for image compression several methods are currendy advocated, including transform coding [Rao90] predictive coding, vector quantization, model based techniques [Aiz95] fractal based techniques [Bar93] adaptive morphological subband decomposition [Egg95] neural networks based approaches [Don95] simulated mid mean filed annealing [Ozc95] and wavelets. Even, when application has only one Figure 1. Throughput Optimization using Simultaneous Algorithm Architecture Matehning and Rephasing computational block, it has been shown that algorithm and architecture selection has very high ....

R.D. Deny, S. Haykin, "Neural Network Approaches to Image Compression", Proc. of the tEEE, Vol. 83, No. 2, pp. 288-303, 1995,

....1. INTRODUCTION Various methods both for lossy (irreversible) and lossless (reversible) image compression are proposed in the literature. The recent advances in the lossy compression techniques include different methods such as vector quantization [1] wavelet coding [2] 3] neural networks [4], and fractal coding [5] Although these methods can achieve high compression ratios (of the order 50:1, or even more) they do not allow to reconstruct exactly the original version of the input data. Lossless compression techniques permit the perfect reconstruction of the original image, but the ....

R. D. Dony and S. Haykin, "Neural networks approaches to image compression," Proc. IEEE, 83, 1995, 288-303.

....(correlated) neighborhoods remain similar across scales, and that this a priori structure can be learned locally from available image samples across scales. Such local information extraction has been prominent in image compression schemes for quite some time as evidenced by JPEG [13] and PCA [14] based approaches, which typically compress the set of nonoverlapping subblocks of an image. Recent compression approaches also exploit the interblock correlation between subblocks [15,16] The goal is to divide the set of subblocks into a finite number of disjoint sets that can individually be ....

R. D. Dony and S. Haykin, "Neural Network Approaches to Image Compression," Proc. IEEE, Vol. 83, No. 2, pp. 288-303, 1995.

....(correlated) neighborhoods remain similar across scales, and that this a priori structure can be learned locally from available image samples across scales. Such local information extraction has been prominent in image compression schemes for quite some time as evidenced by JPEG [14] and PCA based [15] approaches which typically compress the set of 3 nonoverlapping subblocks of an image. Recent compression approaches also exploit the interblock correlation between subblocks [16,17] The goal is to divide the set of subblocks into a finite number of disjoint sets that can individually be ....

R. D. Dony and S. Haykin, "Neural Network Approaches to Image Compression," Proc. IEEE, Vol. 83, No. 2, pp. 288-303, 1995.

....a representative codeword W i for X. Then, only the label (or the index) of this codeword in the codebook is kept to create a compressed image. The main problem in this approach is the design of the codebook. Recently, the use of neural networks for codebook design problem has been investigated [5], particularly Kohonen s Self Organizing Feature Map (SOFM) 6] which is one of the most interesting neural networks for this kind of application. The main interesting properties of SOFM are: self organizing algorithm: it does not need to classify the training image blocks (unsupervised ....

R.D. Dony and S. Haykin, "Neural network approaches to image compression", Proceedings of the IEEE, Vol.83, No.2, February 1995.

....range from repeated optimization with different initialization, and heuristics to obtain good initialization, to heuristic rules for cluster splits and merges, etc. Another approach was to use stochastic gradient techniques [16] particularly in conjunction with self organizing feature maps, e.g. [22] and [107] Nevertheless, there is a substantial margin of gains to be recouped by a methodical, principled attack on the problem as will be demonstrated in this paper for clustering, classification, regression, and other related problems. The observation of annealing processes in physical ....

R. D. Dony and S. Haykin, "Neural network approaches to image compression," Proc. IEEE, vol. 83, pp. 288--303, Feb. 1995.

....(correlated) neighborhoods remain similar across scales, and that this a priori structure can be learned locally from available image samples across scales. Such local information extraction has been prominent in image compression schemes for quite some time as evidenced by JPEG [14] and PCA [15] based approaches which typically compress the set of 3 nonoverlapping subblocks of an image. Recent compression approaches also exploit the interblock correlation between subblocks [16,17] The goal is to divide the set of subblocks into a finite number of disjoint sets that can individually be ....

....images. When a new image is presented, the kernel that best reconstructs each local region is selected automatically and the reconstruction will appear at the output. The division of the signal space in local regions will be implemented via a Vector Quantization (VQ) algorithm as proposed by [15] and many others. Each VQ will be connected to a linear associative memory (LAM) trained to find the best mapping between the low resolution image and the highresolution image, hence capturing the information across scales. In other words, the assumption we make is that the information embodied in ....

R. D. Dony and S. Haykin, "Neural Network Approaches to Image Compression," Proc. IEEE, Vol. 83, No. 2, pp. 288-303, 1995.

....then the samples collected are not sufficient to fully describe the information above f c Hz in the analog signal. This is a limitation of the sampling theorem which we would like to address. It is known that local information in real world images tend to exhibit a high degree of correlation [3,4]. As well, the local structure of various portions of these images (which we define as a local neighborhood minus its sample mean) can also exhibit the same high degree of correlation. We would like to exploit the correlation of such local image information to construct better interpolated points. ....

.... is achieved via vector quantization with a Kohonen Self Organizing Feature Map (SOFM) The SOFM is basically a vector quantizer with two important distinctions: the choice of lattice affects the Voronoi cells created in feature space and a topological ordering of the lattice nodes is achieved [3]. Each of the M quantization (feature) nodes is denoted by F m where m M = 1 2 , # . The Euclidean metric is used for clustering the neighborhoods based on their local structure. The cluster C m is assigned neighborhoods X n p based on their structure as follows: C X F Zvec X F ....

Dony R. D. and Haykin S., "Neural Network Approaches to Image Compression", Proc of IEEE., Vol. 83, No. 2, February 1995, pp. 288-303.

....are called principal components. PCA appears to be involved in some biological processes, e.g. edge segments are principal components and edge segments are among the first features extracted in the primary visual cortex (Hubel and Wiesel, 1962) Mathematically, the KL transform can be written as (Dony and Haykin, 1995): y = Wx (3) where x is an N dimensional input vector, y is an M dimensional output vector (M N ) and W is an M Theta N dimensional transformation matrix. The transformation matrix, W, consists of M rows of the eigenvectors which correspond to the M largest eigenvalues of the sample ....

....x is an N dimensional input vector, y is an M dimensional output vector (M N ) and W is an M Theta N dimensional transformation matrix. The transformation matrix, W, consists of M rows of the eigenvectors which correspond to the M largest eigenvalues of the sample autocovariance matrix, Sigma (Dony and Haykin, 1995): Sigma = Omega xx T ff (4) where represents expectation. The KL transform is used here for comparison with the SOM in the dimensionality reduction of the local image samples. The KL transform is also used in eigenfaces, however in that case it is used on the entire images whereas it is ....

Dony, R. and Haykin, S. (1995), `Neural network approaches to image compression', Proceedings of the IEEE 83(2), 288--303.

....then the samples collected are not sufficient to fully describe the information above f c Hz in the analog signal. This is a limitation of the sampling theorem which we would like to address. It is known that local information in real world images tend to exhibit a high degree of correlation [3,4]. As well, the local structure of various portions of these images (which we define as a local neighborhood minus its sample mean) can also exhibit the same high degree of correlation. We would like to exploit the correlation of such local image information to construct better interpolated points. ....

.... is achieved via vector quantization with a Kohonen Self Organizing Feature Map (SOFM) The SOFM is basically a vector quantizer with two important distinctions: the choice of lattice affects the Voronoi cells created in feature space and a topological ordering of the lattice nodes is achieved [3]. Each of the M quantization nodes (feature nodes or codebook vectors) is denoted by F m where m M = 1 2 , v . The Euclidean metric is used for clustering the neighborhoods based on their local structure. The cluster C m is assigned neighborhoods X n p , based on their structure as follows: ....

R.D. Dony and S. Haykin. "Neural Network Approaches to Image Compression", Proc of IEEE., 83(2):288-303, February 1995.

....It has become increasingly important to most computer networks, as the volume of data traffic has begun to exceed their capacity for transmission. Traditional techniques that have already been identified for data compression include: Predictive coding, Transform coding and Vector Quantization. [15], 19] In brief, predictive coding refers to the decorrelation of similar neighbouring pixels within an image to remove redundancy. Following the removal of redundant data, a more compressed image or signal may be transmitted [15] Transform based compression techniques have also been commonly ....

....include: Predictive coding, Transform coding and Vector Quantization. 15] 19] In brief, predictive coding refers to the decorrelation of similar neighbouring pixels within an image to remove redundancy. Following the removal of redundant data, a more compressed image or signal may be transmitted [15]. Transform based compression techniques have also been commonly employed. These techniques execute transformations on images to produce a set of coefficients. A subset of coefficients is chosen that allows good data representation (minimum distortion) while maintaining an adequate amount of ....

[Article contains additional citation context not shown here]

Dony, R. D., and Haykin, S., (1995), Neural Network Approaches to Image Compression, Proceedings of the IEEE, Vol. 23, No. 2, pp 289-303.

....The resulting hybrid approach has better rate distortion characteristics relative to standard IFS when tested on a standard image. 1. INTRODUCTION A number of new nonlinear techniques for image compression have emerged recently which include fractal methods [1, 2, 3] and neural network methods [4]. They are nonlinear approaches and have been shown to have advantages over standard techniques. However, for fractal coding, the computational requirements for encoding an image are significant. This paper presents a new method which combines features of both approaches, resulting in improved ....

R. D. Dony and S. Haykin, "Neural network approaches to image compression," Proc. IEEE, vol. 83, pp. 288--303, February 1995.

No context found.

R. D. Dony and S. Haykin, Neural network approaches to image compression, Proc. IEEE 83, 288-303 (1995).

No context found.

Dony, R.D. and Haykin, S. (1995b), "Neural Network Approaches to Image Compression," Proc. IEEE, Vol. 83, No. 2, pp. 288-303, February.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC