Image compression at variable bit rates with neural network using dynamical construction algorithm

This paper presents a modular approach of still image compression using dynamically constructive independent node neural networks (DCINNs). A new sub-image block classification technique using wavelet transform and LBG algorithm is proposed for partitioning images into different image clusters. Each module of neural network is trained on a particular image cluster. A modified dynamical construction algorithm is used for building such a network. The DCINN has the inherent capability of producing variable bit rates as it is composed of several independent subnetworks. This feature makes it suitable for transmission of image data over channels having time varying bandwidth characteristic. This architecture is also very robust to hidden node damage.