SAR Image Compression Based on Bandelet Network

Finding efficient geometric representations of images is a central issue in improving the efficiency of image compression. Bandelet provides an efficient way for image representation based on geometric regularity. However, a degeneration of performance will appear in high ratio compression for all the transformation based compression approaches. Moreover, Bandelet is time consuming and lacks of flexibility. In this paper, we construct a Bandelet network for image compression based on discrete Bandelet frame, in which the Bandelet basis is adopted as the activation function in the hidden layer of a feed-forward neural network. The Bandelet basis can provide an efficient representation of image. Moreover, neural networks based methods is more flexible and can achieve high compression ratio with its parallel implementation structure, when compared with transformation based image compression approaches. The construction and the leaning of the Bandelet network are addressed. Experiment results show that it can provide a potential way for SAR image compression.