To attain highly precise prediction through processing much information in a simple constitution by using a neural network for inputting the DC components of a block which operates discrete cosine transformation(DCT), and for outputting the predicted value of the DC components of the DCT of a block to be predicted.

A block to be processed is defined as X, and its DC value is defined as QDCK. Then, the DC values QDCA to QDCD of blocks A-D adjacent to the block X which is being processed for DCT are inputted to an input layer 1 of a neural network 4, and the predicted DC value of the block X which is being processed is obtained from an output layer 3. The neural network 4 is constituted of the input layer 1, intermediate layer 2, and output layer 3, and a back propagation is used as a learning algorithm, and each layer 1-3 is constituted of plural units connected by variable weight. Also, the unit is a transformation system having non-linear input and output characteristics.