A neural network approach to transform image coding

Abstract A neural network approach is presented for transform image coding. It is shown that the three steps in the conventional transform image coding, i.e. the unitary transform of spatial domain image data, the quantization of the transform domain data and the binary coding of the quantized data, can be unified into a one‐step optimization problem. Then, the optimization problem is solved by an appropriately constructed Hopfield neural network whose input is the spatial domain image data and whose output is binary codes. A practical circuit implementation is given to perform the transform image coding. the circuit has rM 2 neurons, where r is the bit‐rate, in bit/pixel, of the coding and M 2 is the size of the images. Each neuron consists of only a non‐linear voltage amplifier, a linear voltage‐controlled current source, a d.c. current source, a linear passive resistor, a linear passive capacitor, and a weighted voltage summer which can be made of a single op amp with some linear passive resistors. Moreover, each neuron is locally connected with no more than b ‐ 1 other neurons by wires, where b is the maximum bit allocated to a transform domain coefficient. Therefore, our proposed approach is particularly suitable for low‐bit‐rate image coding and VLSI implementation. Furthermore, the analogue and parallel nature of our approach matches perfectly the high‐speed requirement of real‐time image coding.