Abstract
This research illustrates the use of a novel model of neural networks, the generalized neural network model, to build connectionist architectures and their processes tailored to the definition of globally and locally adaptive compression systems. This model extends the traditional connectionist ideas to include the behave-act, evolve-learn and behave-control functions of the network, which allow overcoming the drawbacks of previous direct connectionist approaches to the transform image compression problem. Thereafter, the architecture and its processes are used for the definition of a globally and locally adaptive compression system that surpasses known compression algorithms in three main aspects: very high compression rate with a low introduced distortion, ability to tackle a broad set of data and feasibility for on-line real-time compression. The algorithms are backed up by quantitative, empirical and comparative studies to different neural network approaches and known standard compression methods.