Neural network based data compression
Abstract (summary)
This dissertation presents the development, analysis, simulation, and performance evaluation of a data compression-decompression system using a hierarchical multilayer neural network. This system is robust, achieving high compression ratios on large data sets (e.g., images) with small error, short encoding and decoding times, and high generalization. This method does not require generation or use of explicit codebooks, thus reducing complexity and shortening encoding/decoding times as compared to other methods such as vector quantization (1), (2). This data compression-decompression system has been realized by a symmetric four-layer neural network and is trained by using an algorithm referred to as "Nested Training Algorithm" or (NTA) (3), (4). The NTA trains this system in such a way that the network is trained outside-in, (i.e. the outer layers are trained first by using the "Outer Loop Neural Nets" or OLNN, then the inner layers are trained by using the "Inner Loop Neural Nets" or ILNN). The developed data compression system using the NTA divides a scene into several subscenes. During the training phase each subscene is processed independently by the OLNNs. After complete training with the OLNN, the output from each hidden layer of OLNN will be combined and then further processed by the ILNN. By employing the NTA, higher data compression ratios and training efficiency are obtained. System performance has been evaluated by using both synthetic and real world data images. The distortion measures for the reconstructed images are described by the Mean Square Error (MSE), Root Mean Square Error (RMSE), Normalized Mean Square Error (NMSE), and Signal-to-Noise Ratio (SNR) for various degrees of compression. Learning capacity and generalization to other data sets are also evaluated. The results are summarized in tables, graphs, and pictures.
Indexing (details)
Neurology;
Neurosciences
0317: Neurosciences