Neural network technique for image compression

Neural networks (NNs) have been used for image compression for their good performance. However, the image compression convergence time is not efficient. This is due to the fact that the NN is used in image compression and decompression stages in almost all NN methods. The authors propose a method where appropriate NNs are used only at the image decompression stage. The image is decomposed into eight matrices each of which corresponds to values in a bit position. The matrices are saved in reduced form to constitute the compressed image. The NNs are constructed to predict the removed values from the reduced matrices to produce the image in the origin size. This method produces an acceptable and comparable image quality. A compression ratio of up to 81% was achieved while the convergence time can be considered negligible.