Satellite image compression using a bottleneck network

The paper presents a simple neural network suitable for a large size image compression, such as a satellite image. The defined network of N×M×N neurons represents the input, hidden, and output layers respectively. Since M≪N, then this type of structure is referred to as a bottleneck neural network. The new training method and the robustness of the network is satisfied, when this network is trained with images having a small size, and tested with satellite images. The input image is segmented into L blocks each has N elements. Each sub-image presented to the network as input, would appear nearly the same as the output layer. Many experiments were done for different satellite images and the goodness of fit between the original image and the reconstructed image was found to be about nearly 95% with a compression ratio of 4.2:1 even for new images that the network did not learn. It was found that the network is not affected by the geometrical distortions like translation, size, and rotation