By Elwin Chandra Monie
Compression becomes essential for storage and transmission of images since these tasks involve handling of large volumes of data. There is a constant demand for image compression techniques which are faster and more efficient than those that exist already. In the recent years, neural network concepts have been applied to several pattern recognition and image coding problems. Most of these applications are based on the fact that neural networks with their distributed architecture and massively parallel interconnections are fast and efficient in identifying unique patterns in an image.
In this work, a self - organizing neural network called the Adaptive Resonance Theory (ART) network is used in the design and implementation of efficient image compression schemes. The ART network can perform stable clustering of similar patterns of input vectors. Input vectors which are formed from the values of the pixels in a block of image data can be coded into representative vectors called codevectors. The ART network is also adaptive in the sense that it can form new clusters without losing the already stored properties of the earlier patterns. These properties of the ART network are made use of in implementing an Adaptive Vector Quantizer (AVQ). Among the class of ART models, Fuzzy ART network is chosen for designing all the image coding techniques presented in this work as this model is more general and has a simple architecture. First, a simple VQ coder is designed for coding monochrome image data using the ART network. The necessary optimality conditions are proved and it is shown that the prototypes converge to the centroids of the clusters in terms of the similarity measure used. The prototypes can be taken as the codevectors and the nearest codevector to an input is found by the ART network. The advantage of the proposed scheme is that it always guarantees an error performance given in terms of a vigilance parameter r
The major problem in the direct use of AVQ is that the bit rate is high since in addition to the VQ indices, the codebook should also be made available for decoding. To reduce the bit rate, the scheme incorporates a modification using DCT. The codebook that is to be made available is transformed and the transformed codevectors are represented using only a few transform coefficients. Considerable reduction in bit rate is achieved by this technique.
The proposed scheme is purely adaptive and difficulties may occur when used in practical situations. To make it more practical, a method using a universal codebook has been developed. Using this method, the scheme becomes both adaptable and universal. A universal codebook suitable for vector quantizing a wide range of input sources is generated and this is available both in the coder and the decoder. In addition to this, a local codebook that matches the statistics of the image is also generated. This local codebook is of smaller size and it is generated on-line for each image using the proposed scheme, The universal codebook considers the local codebook as the source for vector quantizing it. Only the indices of the universal codebook corresponding to the local codebook are required for decoding instead of the whole local codebook.
In addition to coding of image frames, the adaptive property of the ART network is fully utilized in the coding of image sequences. Each frame is coded by the codebook generated from the previous frame. The codebook that is already generated will get modified when a completely novel input is applied to this scheme. Only the changes in the values of labels and codevectors are required for decoding and hence a very low bit rate is achievable.
The ART scheme is also extended to the coding of color images. Only a few color combinations exist in a color image and these combinations are used as the codebook for vector quantizing the color image. The ART scheme categories similar colors into a small number of categories. The codebook is generated adaptively for each color image achieving efficient vector quantization. This scheme is particularly useful for display of full - color images in a display system with limited color capabilities. Using the ART scheme, the color combinations are limited to a few number instead of a possible eight million. These color combinations are used as a color map to display color images with low visual distortion.
In order to take advantage of the inherent parallelism in the ART architecture, a parallel processing system is more efficient. A hardware scheme is proposed in this work using purely digital circuits to implement image coding techniques described earlier. In this scheme, a systolic arrangement which is a form of pipelining is used. The expressions describing various functions in the ART network are modified to suit easy implementation of the ART algorithm. The image coding method described earlier is also developed based on this scheme. The hardware scheme is decomposed into several cells and all of them work independently. High efficiency and speed are achievable by this method.